🤝｜tech-support

[cosmic furnace]

hm... every start for me took very long to start 🤔

[ornate elk]

The start of the webui or the start of the generation?

[cosmic furnace]

at least theres not so many flux models to use yet and loras for it. if its only read only maybe worth using ssd, but my friend had ssd killed (or smthing went wrong with it) pretty fast using SD on it

well at least it seems like it . It was crucial bx500 1 tb. it didnt even had its warranty ended and started weird behavior... probab not related but still

nah the start of generation was taking too much

[ornate elk]

Crucial ssds are not that great tbh

[cosmic furnace]

😄

y ive read somewhere somewhat reviews telling that...

he shud had taken mx500 not bx500 y?)

lil bit more reliable

[ornate elk]

Yea In the past the bx had problems being the bootdrive in systems

Caused bluescreens

I prefer Samsung, WD and Sandisk

And Kioxia

[cosmic furnace]

yeah mx500 mb still somewhat doubtfull but shud be better )

[ornate elk]

I'm storing all models on an SSD now and didnt got problems yet

[cosmic furnace]

y also from studies made an decision that samsung it is... or kingstons with high wtd

rly? o.0

sd dont kill ssd's? or mb ssds are too new?)

do u have any ssd used with sd for a year or year and a half? hm

[ornate elk]

In my old PC all of Stable diffusion stuff was on a Samsung 860 1tb,
now everything is on a Samsung 990 pro 2tb xD

[ornate elk]

Yes the 860 used it for two years

And now 1 year of the 990

No problems

[cosmic furnace]

wow. ye samsung 860 evo is often advised to buy hm >_>

espec if 22 year or newer

they sasy 21 year have some defect that make some of them broke fast

two years of using sd is wow 🤔

just incase...copied everything from log of forge

when last used

CUDA Using Stream: False - this one is little disturbing...although pony used on forge is running fine & fast -_-

[ornate elk]

Mine 860 evo is from 2018

[ornate elk]

You can add --cuda-stream to the webui-user.bat

It works with ZLUDA

[cosmic furnace]

hmmm... i maybe mistook 860 with 870

ah yeah it was 870 evo that had problems with 21 year release

[ornate elk]

Can you show your forge settings of txt2img?

[cosmic furnace]

how do i run it?

ui ?

[ornate elk]

?

I mean try to generate an image

Also make sure you used the gguf text encoder and and the gguf

[gritty delta]

what does this mean ?

[cosmic furnace]

cuda stream wont hurt?) its better to have it on y?

[gritty delta]

does anyone know where the command line is ?

[cosmic furnace]

webui user bat with notepad

[gritty delta]

icl i don't know what that means xd

this in notpad?

[cosmic furnace]

y

[gritty delta]

and do i just paste the code in a new line ?

[cosmic furnace]

after comand_args like in description

can do a new line after it i guess w/e

[ornate elk]

Test it, but I disnr noticed anything

[cosmic furnace]

@ornate elk u meant this txt generation settings?) not sure if its writes in logs... for now trying to do using a refer image

[ornate elk]

Did you tried to use a lora as model?

[ornate elk]

Never add this please

[cosmic furnace]

shared swap location seems to have even longer generation

[gritty delta]

does it work for batch face swaps?

[ornate elk]

When you do get the error from above?

Can you screenshot the txt2img settings?

And yes with stable diffusion you can face swap

[gritty delta]

its img 2 img

[ornate elk]

Then share that screen

[cosmic furnace]

trying gener with 19000 -_-

[ornate elk]

Try 768x1024 as speed test

[gritty delta]

come into diffuser lounge?

[cosmic furnace]

i have a feeling i might doing smthing wrong with flux... maybe shud watch some guide... but i watched one with fast rewinds felt like more less as simple as auto

[ornate elk]

And make sure to close every programm in the background

[ornate elk]

No I'm on phone, please make a screenshot

And what's your GPU btw

[gritty delta]

1660ti

[ornate elk]

Reactor can't be used like that in img2img, you have to use Extras tab for batch

[ornate elk]

Make sure you have --xformers --medvram --no-half in the webui-user.bat

For the best performance for this card

[cosmic furnace]

seems like generation warming up again is very long =\

[ornate elk]

Check your taskmanager

There look at RAM and vram usage

[cosmic furnace]

y pc feels overloaded with something

even stuttering sometimes lil bit & long opening taskman

hdd 100% load memory 69% cpu-gpu jumps from 0 to 3%

[ornate elk]

Im AFK now

[cosmic furnace]

okay 😦

[gritty delta]

got this error

[cosmic furnace]

again blurry image o.0

and took somewhere 2.5 mins to gen a 1024-768 -___-

too slow for 1024-768 =\

2nd try not so blurry increased gpu used from 19000 to 19400 and it ended up making image longer 3m-17s ... indeed. Weird +)

any advices about setting & setting up flux? 2-30 still slow for a 1024-768 y ? 🤔

on pony auto makes 1024-1536 like in 10-12 sec -____-

[fierce hollow]

Friends, I am glad to have questions about SD.
The question concerning all SD models is What is the problem to create a model for cruelty (severed limbs and so on, is this the censorship of the site or are there other problems?)
The question regarding Flux is what is the problem or again any prohibitions on creating Flux NSFW|| (Full version with pron)||

[ruby yoke]

you don't need a model for that, just drop your negative prompt terms into the positive prompt - if you're on your own machine instead of someone's website service that filters stuff like that out, you'll get horror shows

[fierce hollow]

I will try... And with regard to the second question?

[ruby yoke]

i sort of answered both questions at the same time - if you're on your own machine, and you're not using the API, just using one of the open source models, then the 'prohibition' is what you decide - and what your local laws state as legal for what you can do on your own computer in yoru own home.

if you're using the API - you're under the TOS from black forest for what you can do

if you're on someone's website, you're dealing with what they allow

[fair oxide]

Hi friends, I have a pic that I generated in stable diffusion using epicrealism_naturalS model. I loved the face/character it generated. However stable diffusion isn't generated similar images in the img2img tab when I'm passing in the original image. I didn't even change the prompt, just want different variations of the same face. any help here? Been struggling with this for sometime now...

[ruby yoke]

that's not what img2img is for. the AI uses your uploaded image as a prompt and will use the color pallet, and shapes but will then fill in those areas with your text prompt. what you want is something like this https://huggingface.co/spaces/jbilcke-hf/FacePoke

[fair oxide]

Thanks for answering!

I checked what you shared, that's not what I want. I want similar character face but different pics (poses, attire, angles) etc.

[ruby yoke]

okay well - my only suggestion is to use facepoke to get the expessions, then generate the images you want, and face swap in the facepoked faces

[fair oxide]

🤔. Theoretically that makes sense. But I'm not sure the type of expressions I'd want. I want the model to surprise me and be creative. Same with the body/background. So I was hoping to give it a sample image and it can create similar face body but in other expressions, poses, attire, angles.

[ruby yoke]

that's not how it works. there are tools that do that, but not stable diffusion itself. you could try putting a controlnet workflow together in comfyUI with some open pose nodes, a face swap node, some other stuff, and that might work. but that's not point and click

[fair oxide]

that seem too much setup for my liking and skills.

so there's no way in sd to say generate similar images?...

[ruby yoke]

stable diffusion is an image generator. to get it to do what you want isn't going to be a point and click solution. you'll either have to create a workflow with all the various functions necessary or use a pre built tool - and the ones i've seen that aren't for things like iPhones or androids are mostly for video, not still images

[pale magnet]

hey guys, good night. sorry for the constant messages, but i'm still in need of help. it's for my final paper at college so it would be really appreciated. i've tried running some lora notebooks but they don't seem to work properly, and i'm not sure what i can do. i needed to finetune a model in a small dataset of mine (about 300 pictures). i've tried autotrain dreambooth as well but it didnt work

[carmine ibex]

Hi, I'm trying to install torchvision because I'm getting "ERROR: Could not find a version that satisfies the requirement torch==2.3.0+cu121". I already tried version torchvision-0.18.0+cu121-cp311-cp311-win_amd64 but I don't understand why I keep getting that message. I thought it would be compatible.

[ornate elk]

That can work with controlnet and IP-Adapter

[ornate elk]

Hey, which webui you use?

[fair oxide]

I tried but not working, at least for me. Any sample? I can also DM if that helps to explain the situation better.

[ornate elk]

Okay, have you tried Reactor for Faceswap?

I'm not at the PC to show the Controlnet IP-Adapter settings

[fair oxide]

never heard of it. let me try

[gritty delta]

Hey I'm having issues when batch face swapping on SD, using reactor.

after the face swap the face is really blurred / bad quality compared to the rest of the image.

here's and example -

does anyone have a fix to this ?

settings

[ornate elk]

Try enable the upscale by and dont use the resize by.
Then try the upscaler IJ reactor

[gritty delta]

apparently reactor has a 128x128 cap to prevent deepfakes i've been told

so all images will have a blurred / low quality face

[valid gorge]

Is there a way to somehow create images in batches, if I don't have the empty latent node (or similar) and I'm doing inpainting in Comfy?

[ornate elk]

Hmm then you can use the standalone tool called
Roop-unleashed

I will so some reactor tests later

[gritty delta]

does that still work ? , thats what i tried in the first place.

but it had so many issues with things being out of date or smth

[gritty delta]

or this is diffrent from roop?

[ornate elk]

Its based on roop but further advanced

[gritty delta]

i see

i will try to install now and see how it goes

[ornate elk]

The roop extension is deprecated thats why there is Roop-unleashed as standalone tool

And Reactor and facefusion

[gritty delta]

is says he after running the "windows_run.bat" installation,

to run roop unleahsed you just click the bat file again.

but when i run it i get this error

[vocal burrow]

1/ dont install stuff on the desktop
2/ which python version are you using ?

[gritty delta]

3.13

[visual bane]

I am trying to generate the interior with a custom model of sdx 1.5. I changed the controlnet to normalmap and mlsd, but the interiors are coming out small and the results are not as expected. The custom model is for interior design. Can anyone suggest improvements I can perform?

Choose random theme and make interior design for this bedroom with large furniture
Negative prompt: A distorted, low-quality room with altered walls, ceiling, and unrealistic doors and windows. Blurry details, bad lighting, unfinished areas, and mismatched elements create a chaotic, sketchy feel. Distorted furniture, oversized decorations, and fake shadows add to the bad perspective and clutter, giving it a surreal, cartoonish look.
Steps: 20, Sampler: Euler a, Schedule type: Automatic, CFG scale: 13, Seed: 1105439493, Size: 626x351, Model hash: aae64db561, Model: model, Denoising strength: 0.7, ControlNet 0: "Module: mlsd, Model: mls [a78de37a], Weight: 0.3, Resize Mode: Crop and Resize, Processor Res: 720, Threshold A: 0.15, Threshold B: 0.3, Guidance Start: 0.0, Guidance End: 1.0, Pixel Perfect: False, Control Mode: Balanced", ControlNet 1: "Module: normal_midas, Model: nomral [3e0bc3c3], Weight: 0.45, Resize Mode: Crop and Resize, Processor Res: 512, Threshold A: 0.4, Threshold B: 0.5, Guidance Start: 0.0, Guidance End: 1.0, Pixel Perfect: False, Control Mode: Balanced", Version: v1.10.1

[vocal burrow]

I would stick to the usual and known to be compatible one, 3.10.11 just to be safe

and still, don t install stuff on your dekstop

[gritty delta]

hmm i see , soo should i reinstall this or how do i fix it so its in a folder or not on the desktop 🤣

[vocal burrow]

Reinstall it somewhere else. It s not like there is something worth saving at the moment anyways.

[gritty delta]

so i installed it inside a folder and still getting this error

[vocal burrow]

more of a question for #📝｜prompting-help
But I don't see that much problem with the output technically speaking. Also "sdx 1.5" is not a thing, sdxl is, sd1.5 also is but not sdx 1.5 nor sdxl 1.5. And sd1.5 and sdxl kinda require different prompting style to get optimal output.

[vocal burrow]

still in desktop

[gritty delta]

im so lost bro

[gritty delta]

its not on desktop anymore and still the same

how do i make sure that its in the directorty

[vocal burrow]

what s in your d:\roop-unleashed folder ?

[gritty delta]

this installs when i run the bat file

[vocal burrow]

Then yes the computer does not lie... there should be a installer.py file in there (and macos installer too).

How did you "get" roop unleashed ?

Maybe your antivirus deleted it automatically.

[gritty delta]

I think i found the issue

now just waiting for all the packages etc to install

[gritty delta]

hey man, So i've figured it all out.

it was an issue with the folder i was downloading from i just had to redownload the initial folder from github and then just run the installer again.

but now the main issue i was running into previously.

the images are still processing with the blur on the face idk why,

its like every face swap is censored

[rose siren]

Hey, new to the tech. Wanted to now what models you'd recommend which could run on a RTX 2060 with 6GB VRAM. I don't need super good quality, just something to test out along with ControlNet (canny edge).

[carmine ibex]

I managed to fix it. Thanks though

[gritty delta]

nvm fixed

[rose siren]

Would anyone happen to know how AUTOMATIC1111 / Forge can produce Pytorch models from just the safe tensors. Try to read source code but load_model function seems to have a lot of moving parts.

https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/82a973c04367123ae98bd9abdf80d9eda9b910e2/modules/sd_models.py#L786

[kindred pivot]

Hey there, hope everyone is doing fine. Anyone is experienced in inpainting in comfyui ? I'm struggling a lot. I want to inpaint some stuff in the plain beige part of this image. Any idea how ?

[ruby yoke]

put in a prompt, select the area you want to inpaint on, and generate. the AI will target the area you selected for generating your prompt into

[rocky dragon]

What would be the most complete way to transfer an entire A1111 WebUI installation (settings, models, LORAs) between computers?

Is there a guide for this?

I'm running out of space and need to move my installation to a different drive

[ornate elk]

so you want to move the webui to an other pc or only to an other drive on the same pc ?

for the PC option: Clean install A1111 on the new machine using the install guides. Then copy the models, outputs, extensions, embeddings folder over and your done.
For the other drive on the same PC option:
move just the models folder (because its the biggest part) to any other drive and in the webui-user.bat add --models-dir "D:\PathTo\TheOtherDrive"

[hushed niche]

hey guys, im very new to this, an i have a question about running loras on forge, if theres no activation word mentioned in the lora i dw, is there any way to actually use it? i get going to the lora tab and clicking on, then it adds it to the prompet, but it makes 0 difference in the pics i generated, am i doing something wrong? or is it just not possible and they are only working with comfyui?

[ornate elk]

hey, make sure the lora is compatible with the model version

[tawdry vault]

when I add in multiple loras to my prompts I always get this error: Lora not found: WesternCartoonClassicDisney100, andav, shan

Anyone deal with this?

[ornate elk]

Which webui do you use?

[tawdry vault]

AUTOMATIC1111 Stable Diffusion web UI

[ornate elk]

And the loras are all in models/lora ?

And compatible with the model?

[tawdry vault]

Yeah I have tested all of the loras, they work on their own fine. I just can't add them onto the prompting, I will always get that error under the image after generation

Is this the wrong way to add on additional loras?

score_9, score_8_up, score_7_up, solo, (front view, centered, symmetrical:1.2), portrait, 1man, middle-aged sheriff, brown hair, stubble, brown eyes, looking at viewer, cartoonish lora:WesternCartoonClassicDisney100:0.4, lora:andav:0.35, lora:shan:0.35

the names are correct

[ornate elk]

Looks fine

Can you show the full cmd log when you get the error?

[tawdry vault]

There is no error in the command log, but here is what the output looks like on the UI after image generation

score_9, score_8_up, score_7_up, solo, (front view, centered, symmetrical:1.2), portrait, 1man, middle-aged sheriff, brown hair, stubble, brown eyes, looking at viewer, cartoonish lora:WesternCartoonClassicDisney100:0.4, lora:andav:0.35, lora:shan:0.35
Negative prompt: score_3, score_4, score_5, censored, 3d, monochrome, female, woman, 1girl, athletic, muscular, frame, border, profile, angled, side view, big eyes, large eyes
Steps: 10, Sampler: DPM++ 2M, Schedule type: Karras, CFG scale: 7, Seed: 1192914656, Size: 512x512, Model hash: 67ab2fd8ec, Model: ponyDiffusionV6XL_v6StartWithThisOne, Version: v1.10.1

Lora not found: WesternCartoonClassicDisney100, andav, shan

[ornate elk]

Do you have any lora extensions installed?

Or lycoris

[tawdry vault]

No

all I have installed extra is .safetensors

[ornate elk]

Can you still show the full cmd log? Maybe there is some hint

Or did you messed with the Settings of the webui, then you could reset them by deleting the config.json ans the ui-config.json

[tawdry vault]

sure here is the command log

[tawdry vault]

I have not touched the settings much but I'll try that

[ornate elk]

There are a view errors and this is weird:
Reusing loaded model WesternCartoonClassicDisney100.safetensors [7b0e144b36] to load
Is western even a lora ? (Smaller than 2gb)

Also what's your gpu? Your using CPU only mode.

[tawdry vault]

I don't have gpu I have integrated graphics on my laptop lol so I run on CPU

[ornate elk]

Ah okay

[tawdry vault]

https://civitai.com/models/305625/western-cartoon-classic-disney-pony-diffusion

it's this

[ornate elk]

Thats a lora

Check your models/stable-diffusion folder and delete it if its in there

It has to be only In models/lora

[tawdry vault]

oh that is where I have put all of my stuff

[ornate elk]

Ohh that's the issue then xD

[tawdry vault]

all of this goes to models/lora?

[ornate elk]

In models/stable-diffusion go only files that are 2gb or larger

The shan.pt file is an Embedding and goes into stable-diffusion-webui/embeddings

The rest (smaller than 2gb) goes into models/lora

And to get a bit more speed with CPU you can add
--opt-sub-quad-attention
to the webui-user.bat

[tawdry vault]

thank you

[ornate elk]

No problem

The loras should now show up in the lora tab

[tawdry vault]

I'll confirm in a moment

[ornate elk]

I'm off now cya

[river pollen]

For some reason I don't have new tabs appearing from extensions (i.e. from TemporalKit) in my forge ui. Is it a common problem? All the tabs I have available can be seen on the screenshot

[cunning robin]

hey uhm, why it stopped working even Pony?

[ornate elk]

Thats a our of vram error

So could be a to large resolution or wrong upscale settings

[cunning robin]

but it worked before, why it stopped working?

[ornate elk]

Maybe you changed a setting? Idk ^^ can you show the txt2img settings when you get the error?

[ember pine]

installed this fork GitHub - lshqqytiger/stable-diffusion-webui-amdgpu-forge: Forge for stable-diffusion-webui-amdgpu (formerly stable-diffusion-webui-directml)
my cpu 5800x3d and gpu 6800xt install rocm from amd and python and git
user bat file just use zluda and auto launch
i got it running but my problem is that its using my cpu to generate images instead of my gpu... any help would be appreciated it

[cunning robin]

ah yeah well the prompt is definitely bigger

do i send it via dm? positive and negative?

[vast pulsar]

Hey guys i'm new to comfy so i'm kinda lost, what does this mean?

[ornate elk]

Sure, the prompt isnt needed only the settings and the model

[ornate elk]

Hey, this is a hip bug as it found your CPUs iGPU (gfx1036) before your dedicated GPU.

To fix this you have to open up the Device Manager and under Display Adapters disable the Radeon TM Graphics.

Or you can add: set HIP_VISIBLE_DEVICES=1
to the webui-user.bat

Then relaunch the webui-user.bat

[ember pine]

under display adapter its showing amd radeon rx 6800xt

i just chated with gemini lol but yea because i saw my gpu utilization low when make an image but the cpu usage going to 22% i thought it was using my cpu not gpu but then looked again the dedicated gpu memory is 100% being used when testing with Hires its weird cuz usally when i game i just at utilization % lol

[ornate elk]

Ah okay so it gets used normaly

[ember pine]

is that normal gpu utilization 2% vram 100%

[ornate elk]

When using hires make sure to always set the hires steps to 10

[ornate elk]

What's in your webui-user.bat?

100% is a bit much

Or show your txt2img settings

[ember pine]

--use-zluda --autolaunch --opt-sub-quad-attention --no-half

the last 2 just threw them in dont know what for XD

[ornate elk]

Yea dont use the last two

Instead add --skip-ort

[ember pine]

i got a fork of a1111 running it was doing 22it/s i delete it cuz i didnt see gpu % go up and was seeing cpu 22%

but with this forge webui im getting like 5it/s

[ornate elk]

So your using forge ?

[ember pine]

yeah

amd fork

[ornate elk]

Ah okay

Because In forge its not --use-zluda its --zluda

[ember pine]

ok let me try that and --skip-ort

[ornate elk]

--skip-ort is not needed in forge

That was for Auto1111

But would recommend using Auto1111 over forge for AMD

I'm using forge only for flux and Auto1111 for the rest

[ember pine]

i tried but when loading it, it would say no rocm agent found

[ornate elk]

That could mean your rocm hip sdk isnt added in path correctly

Did you installed hip SDK 6.1 ?

[ember pine]

no i installed 5.7

and its showing in system variables

[ornate elk]

Oh 5.7 for 6800 is bad

6.1 works much better

[ember pine]

the zluda notes on github said use 5.7 and 6.1 they havent 100% ...

[ornate elk]

Idk which guide you followed but it seems old then

[ember pine]

ROCm Toolkit 5.7 was found.
fatal: No names found, cannot describe anything.
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: f2.0.1v1.10.1-1.10.1
Commit hash: 93211db60a93840782454cf6f54a62bdc088a7bc
Using ZLUDA in C:\sdk\stable-diffusion-webui-amdgpu-forge.zluda
WARNING: no ROCm agent was found!
Launching Web UI with arguments: --zluda --autolaunch
Total VRAM 16368 MB, total RAM 32693 MB
pytorch version: 2.3.1+cu118
Set vram state to: NORMAL_VRAM
Device: cuda:0 AMD Radeon RX 6800 XT [ZLUDA] : native
VAE dtype preferences: [torch.bfloat16, torch.float32] -> torch.bfloat16
CUDA Using Stream: False

yeah even forge is no ROCm agent found

[ornate elk]

If you want to do everything correctly, checkout my AMD Guides.
I made them for each webui and even lora training.
You can find them here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

Feel free to ask if any step isnt clear

[ember pine]

ok ty

[ornate elk]

But uninstall hip SDK 5.7 and restart the PC before installing hip SDK 6.1

[ember pine]

i think its working im getting 22it/s

thank you so much your amd guide did the trick

[ornate elk]

Perfect, no problem 🙂

[ember pine]

check your dms

[gritty delta]

does anyone know what this does on roop unleased ?

also quick question if anyone knows how to optimise face swap quality, msg me ❣️

[night walrus]

Will any 3.5 variants be trainable on 12GB of VRAM?

[fierce hollow]

How is the SD 3.5 model better than flux? Does Forge - sd 3.5 support it?

[solemn moon]

Bros
Will 3.5 need all new control nets etc

[midnight roost]

hi i have finally made stable diffusion to work i have a 6950 xt gpu and 7800x3d cpu and i was wondering is it not possible to make my cpu also do some of the work?

[orchid comet]

Anyone know if we still need this node for 3.5?

[quaint sleet]

what are the minimum specification requirements for SD3.5?

[ornate elk]

Hey, nope the CPU only will do some tasks that the GPU can't do. For example Faceswap with reactor on amd.
And did you used zluda and hip SDK 6.1 ?

[normal fox]

seconded

[errant tundra]

I'm getting this error and idk why, happens on stock apply controlnet too

btw everything is up to date according to the manager.

[gentle sundial]

Hi I need help please,
I'm getting this error
Stable diffusion model failed to load :((

it was working fine yesterday

@ornate elk 😦

[errant tundra]

what are you using

[gentle sundial]

you mean what model?

Pony

[errant tundra]

well that but also the program you are using. I am only experienced with Comfy cause it's so versatile

[gentle sundial]

Oh no not comfy 😦

[errant tundra]

sorry can't help you then as I use comfy due to having more capability and versatility.

[gentle sundial]

you need high Vram to use comfy right?

[errant tundra]

no not particularly. I use less vram with comfy than with other stuff

it has an option, --lowvram that will keep it stable but just run a bit slower if you don't have the vram

[gentle sundial]

is it hard to install?

[errant tundra]

no pretty easy

just has some learning curve

[gentle sundial]

yeah I know that already cause I'm a game dev

do you have link to how can I install it ?

[errant tundra]

You'll like comfy's node system

just google it.

once you get comfy installed you should install the comfyUI manager

which will make it easier to update, download new custom nodes, download models, etc

the manager simplifies everything

[ornate elk]

Can you share more information?

Like a cmd log

Or txt2img settings

Or your GPU

[gentle sundial]

yes sir one sec

[errant tundra]

once his problem is resolved could I get some help as well

[gentle sundial]

will it replace everything or would it installs comfy separately?

[errant tundra]

it's seperate completely

it's a replacement for the webui

[gentle sundial]

so I wont be able to use webui anymore?

@ornate elk

[errant tundra]

it's 1 or the other. WebUI is more simplistic and user friendly while comfyUI is more in-depth and advanced

[gentle sundial]

they both give very close resuilts don't they ?

[ornate elk]

the full cmd log pls

[errant tundra]

ehh. depends. you can accomplish more with comfyUI.

this is my master workflow for comfyui. Something I have been working on for several months

yours likely will not ever get this complicated

this is the default workflow for text2img

it can be as simple or as complex as you'd like

[gentle sundial]

here it is I hope I did it correctly

[gentle sundial]

damn

[gentle sundial]

Thanks for info!

[ornate elk]

okay, first go into the extensions folder and there delete the roop extension. Its broken as its not updated since over a year. it causes issues.
Also the CloneCleaner extensions is not updated since a year so maybe delete that too.
Then delete the venv folder.
Then relaunch the webui-user.bat and try to load the model again

if you only have 16gb RAM you need to increase oyur windows pagefile

[errant tundra]

Could I possibly get some help as well

[gentle sundial]

Thank you very much! trying now

[ornate elk]

cant help much with comfy, but make sure its updated and the custom nodes too.
Then try a normal single controlnet to verify if that works

[errant tundra]

yea that's the problem is even normal cn isn't working and everything is up to date according to the manager

[vast pulsar]

So guys, webui is much slower than comfy for me since i comfy has the ability to launch "with nvidia gpu" any way i can make my webui faster?

[ornate elk]

whats in your webui-user.bat?

[vast pulsar]

pretty sure it's still the default

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=

call webui.bat

was kinda expecting a 3070 to perform a lil better lol

[ornate elk]

your missing the performance args, add:
--xformers --medvram-sdxl --no-half-vae to the commandline_args=
then save and relaunch

[gentle sundial]

Just out of curiosity , how long it take you to render high quality image?

[vast pulsar]

that's already in another bat

[vast pulsar]

depends, in comfy? seconds

[ornate elk]

another ?

[vast pulsar]

in webui? pfff if i'm chnaging loras then maybe like

30 mins

[ornate elk]

you should only have one

[gentle sundial]

damn!

[vast pulsar]

webui.bat

set COMMANDLINE_ARGS=--precision full --no-half --lowvram --always-batch-cond-uncond --opt-split-attention

[ornate elk]

oh thats bad

[vast pulsar]

also i've tried medvram despite making my pc not shit itself it also makes generating even slower so i went back to low

[ornate elk]

remove them and only add: --xformers --medvram-sdxl --no-half-vae

you massivly slowed yourself down with the ones you used

[vast pulsar]

well it's the default, i'm way too scared to change hem alone lmao

[midnight roost]

zluda but 5.7 hip

[vast pulsar]

but yea i'll try those

[ornate elk]

ah okay, would suggest hip sdk 6.1 for a 6800 and 6800xt because its native supported by rocm hip

much better for the future use

[midnight roost]

ok can i just install it or is it something fancy i need to do? i def dont wanna break anything

[ornate elk]

you cant break much, and you always can look up my AMD install guides if something breaks.
For example yesterday i added How to upgrade hip and zluda on my troubleshoot section.
You can find my AMD Guides here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides
The troubleshoot is at the bottom

[vast pulsar]

oh wow the booting was defenitly faster

[little gate]

Curious, how does SD 3.5 interact with the webui forge?

[midnight roost]

ok thx what the reason to update it btw?

[ornate elk]

Better performance, and better support for programms like LLM software. Also rocm 5.7 will run out of support anytime.
Also ZLUDA development is only focusing on 6.1 now

[gentle sundial]

is automatic forge the new version of 1111?

[ornate elk]

its just a different webui based on auto1111

also did your model loaded now?

[vast pulsar]

HEY

IT WORKED

[gentle sundial]

still installing alot of stuff

[gentle sundial]

what's the difference between them?

[midnight roost]

btw i did the edit variable pathing thing in windows do i need to redo that step? after update

[ornate elk]

you need to check the variable and if it points to 5.7 change the 5.7 to 6.1

if you only have %HIP_PATH%bin in there its fine

[midnight roost]

and should uninstall hip before updating?

[ornate elk]

the old one yes

[midnight roost]

oh ok

[vast pulsar]

damn bro thx sm, from 30 min gens to seconds

i was going crazy cuz every time i changed lora here we go 30 mins

[errant tundra]

still having issues with ControlNet on ComfyUI

[ornate elk]

@midnight roost if you didnt found the troubleshoot section, here is the to do:

How to upgrade or downgrade ROCm and ZLUDA:

1. Uninstall all HIP SDK programms in the System control.
2. Download the HIP SDK Verison you want to switch to and leave everything at default settings.
3. Reboot the PC.
4. Delete the content of the C:\ZLUDA folder so that its empty.
5. Download the correct ZLUDA Version depending on your installed HIP SDK Version.
   https://github.com/lshqqytiger/ZLUDA/releases/
6. Extract the files into the C:\ZLUDA folder.
7. There make copy of the cublas.dll and the cusparse.dll and nvrtc.dll repaste them inside the folder.
8. Rename the copies to cublas64_11.dll and **cusparse64_11.dll **and nvrtc64_112_0.dll
9. You can use these 3 files to replace them in torch folders of different webuis.
10. Check the PATH in System Environment Variables. Make sure only the** %HIP_PATH%bin** is listed there or the correct path to the HIP SDK ROCm Version you installed.
11. Delete the venv and the .zluda folder if a Webui has it and relaunch the Webui.

[ornate elk]

no problem ^^

[gentle sundial]

Thanks it worked, but I'm getting this error now
RuntimeError: "fill_cpu" not implemented for 'Float8_e4m3fn

[errant tundra]

figured it out. Apparently these are LLLite ControlNet Specific and I didn't have the proper nodes

[ornate elk]

can you try load an other model?

[ornate elk]

forge currently doesnt suport 3.5 i guess

[little gate]

oh

[gentle sundial]

yeah it worked ,thank you! , should I delete and redownload the model? also since you said delete clone cleaner and roop, is there any thing else I can use to get some variations for characters look , I have no use for face swap

[ornate elk]

hey np, for variations you can try using wildcards and the dynamic prompt extension

https://www.youtube.com/watch?v=bQK5diN59NA

and yes you can delete the model that wont work and try redownloading it

[median geode]

Hello, may i ask which type of hard drive is more suitable for installing Stable Diffusion. Is there a significant difference between a mechanical hard drive and an M.2 solid-state drive (SSD)? How much space should generally be reserved for storing models to complete the work?

[ornate elk]

hey, the best is to install the webuis and ai tools on an SSD. the models can be stored on an ssd too (preferable) but not need if you dont have much space. If you have a hdd it should have 7200rpm. that will work good too

what is important, that the windows pagefile is stored on the SSD

if not, that can slow down the whole OS when the RAM gets filled to max

[gentle sundial]

thank you and for pony and xl wild cards don't work

[ornate elk]

they should work normaly

wildcards are not based on models

[gentle sundial]

hmm I iwll try them now thanks!

[errant tundra]

Hey could I bother you for a sec about this error I am getting when trying to use controlnet?

[main nest]

yo guys is ther ea guide how to start with the new SD 3.5

[median geode]

How much space would be suitable for models, like 500g?

[ornate elk]

yea, for sure. but it depends on you, if oyu find a few models you like you dont need that much

1.5 models are 2-4gb
SDXL and Pony models are 6gb
While Flux Models can be from 8 up to 16gb or even more

[ornate elk]

that means your using 2 files that dont match like a lora made for 1.5 and an sdxl model

or a controlnet model made for sdxl with an 1.5 model

[median geode]

What impact does the type of hard drive have tho

[ornate elk]

the image generation isnt affected by the drive where the models are stored on.
Only the Model loading time will be longer.

that isnt impactful but if the drive is slow, it can take very long with sdxl/pony, sd3 and flux to load or switch to

i havent had any loading problems on my old 4tb 7200rpm hdd

but it will be much slower on a 5600rpm drive and even slower on an external usb hdd

[median geode]

Thank you fr

[ornate elk]

no problem 🙂

[night vapor]

Hi, I don't normally use github or hugging face. How do I download 3.5 Large to my M3 macbook? I tried looking at the guides in pinned but I don't think they're for the latest model.

[gentle schooner]

"The size of tensor a (128) must match the size of tensor b (2) at non-singleton dimension 6"

[ornate elk]

hey, you can just download the sd3.5_lare.safetensors file from huggingface:
https://huggingface.co/stabilityai/stable-diffusion-3.5-large/tree/main

[gentle schooner]

Here is the workflow for the error

[ornate elk]

@gentle schooner the screenshot is to low res

but check the model and the controlnet model, if they are compatible

[night vapor]

would it be deploy? I'm not a programmer by any means

[gentle schooner]

any idea?

[wraith jackal]

Hello, I'm trying to use Automatic1111's Web UI, I just installed today and I'm encountering some technical difficulties

[tawdry galleon]

Are the Clips & TE's for 3.5 large, and 3.5 large turbo the same/compatible with each other? Meaning will one set work for both models?..... or are they separate and only work for the specified model?
Just wondering as they are all identical in size?

[carmine ibex]

WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
PyTorch 2.4.0+cu118 with CUDA 1108 (you have 2.4.0+cu121)
Python 3.11.9 (you have 3.11.9)

Any idea which xformers version I am meant to install?

[ornate elk]

Do you have a webui installed already on Mac?

[ornate elk]

Hey, is the model you try to use 2gb or larger?

If not, its not a model

[wraith jackal]

I'm not sure

https://civitai.com/models/393824?modelVersionId=512724

I just want to generate general anime content really...

can you guide me? :3

[ornate elk]

Thats a lora, not a model. It goes into models/lora

[ornate elk]

It has to be uses together (on top) of a model

[wraith jackal]

is there a way to not use model and just generate a character with descriptions?

[ornate elk]

Nope, to generate an image you always need a model (checkpoint)

[wraith jackal]

https://civitai.com/models/34511/vtuber-or-or-nakiri-ayame

this a model?

[ornate elk]

Nope, check the info box, it says Lora

This is a model:
https://civitai.com/models/269232/aam-xl-anime-mix

[wraith jackal]

damn

[wraith jackal]

thanks, I'll try to use it

[ornate elk]

Np, put that into the models/stable-diffusion folder and move everything smaller than 2gb to models/lora

[gentle schooner]

@ornate elk Could you help me perhaps? I'm stuck on it for a great time

i think that it's an incompatibility problem but i'm not sure at all

[ornate elk]

Yea the Controlnet model seems incompatibel, try this:
https://huggingface.co/raulc0399/flux_dev_openpose_controlnet/tree/main

[gentle schooner]

ty i'll try it rn

[ornate elk]

Which webui do you use?

[wraith jackal]

I got

OutOfMemoryError: CUDA out of memory. Tried to allocate 20.00 MiB. GPU 0 has a total capacty of 4.00 GiB of which 0 bytes is free. Of the allocated memory 4.72 GiB is allocated by PyTorch, and 213.64 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[night vapor]

no

[ornate elk]

What's your GPU?

And what's in your webui-user.bat?

[wraith jackal]

git pull
@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --xformers --theme dark --api

call webui.bat

[ornate elk]

add --medvram

[ornate elk]

And then in the webui settings enable FP8 mode for sdxl. Then hit apply and reload

[wraith jackal]

sorry, but where can I enable the fp8 model for sdxl?

maybe in settings?

[ornate elk]

To use stable diffusion, and the models like sdxl, 3.5 etc you need to have a local webui installed.
In the pinned messages you'll find All in One install script for MAC,
with the webui auto1111 or forge you can generate images and use different models.
SD 3.5 currently only works with ComfyUI which I dont have a guide rn for Mac, but Forge Webui will support it pretty soon too I guess.

[ornate elk]

On the left site is a search box. Type fp8

Then enable it only for sdxl

Then hit apply and reload ui

[wraith jackal]

thanks

[carmine ibex]

Comfyui

[wraith jackal]

I hit apply

then I pressed reload ui so I'll see how it works

OutOfMemoryError: CUDA out of memory. Tried to allocate 14.00 MiB. GPU 0 has a total capacty of 4.00 GiB of which 0 bytes is free. Of the allocated memory 5.20 GiB is allocated by PyTorch, and 162.50 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[main nest]

?

[ornate elk]

Okay then check this thread:
https://github.com/facebookresearch/xformers/issues/1079
It seems there are some dev versions and a link to all xformers version you can try.

[ornate elk]

Currently only comfyui supports SD 3.5

So set up comfyui, check the comfyui examples page:
https://comfyanonymous.github.io/ComfyUI_examples/sd3/

[ornate elk]

Try close the cmd and webui and relaunch the webui-user.bat

[wraith jackal]

some tutorials to fix errors, I see they manipulate a file that has import torch but idk which file it is

[main nest]

well im down if I can create ralistic looking models such as this one or even better

[ornate elk]

Please don't touch any files or edit stuff other than the webui-user.bat.
Your error is just a normal "not enough video memory" error

What you can try if you still get the error is to change --medvram to --lowvram

And then relaunch

If that still wont work with the sdxl model, you can try using the forge webui which could be better for your GPU.
Or use smaller models like 1.5 based ones

[main nest]

well do you have a guide how to set up a comfyui

I remember you had for SD

which was pretty simple

[ornate elk]

Yea the first link in the pinned messages

All setup guides are there

[main nest]

Idk but you are a goldfish

in the good way

[old parrot]

How can I prevent this while using XL on 6gb 1660S?

[wraith jackal]

Connection errored out.

[main nest]

quick question, whats your RAM

[ornate elk]

How much ram do you have?

Also you installed the webui the wrong way

(By downloading it as zip and placing it in a OneDrive synced folder)

[ornate elk]

Have you enabled fp8 mode for sdxl ?

Or which webui do you use?

[gentle schooner]

Do you have the flux shnell version? It seems to be the flux dev one

[old parrot]

Nope I'm using Forge

--unet-in-fp8-e4m3fn this?

[ornate elk]

I dont think there is any model for schnell specifically

[ornate elk]

Hmm in forge idk if its doable.

[gentle schooner]

well it gives me this error

so i don't believe that it works with the shnel version one

[ornate elk]

Ah that error is just a node error

The node doesn't support other Controlnet models it seems

[wraith jackal]

8.00 GB (7.34 GB usable) @main nest

[old parrot]

Any suggestions on commandline args in general for the 1660S on forge?

[wraith jackal]

it's too little?

[ornate elk]

Its to less for sdxl. You need to increase your Windows Pagefile to make sdxl models work

[main nest]

depends what you want from the SD

[wraith jackal]

how I do that?

[wraith jackal]

tbh I just want to create veadotube model for a programming project

[ornate elk]

@wraith jackal download this model and try it:
https://civitai.com/models/84586/aam-anylora-anime-mix-anime-screencap-style-model
Its the 1.5 version, it will work

[wraith jackal]

I want to be able to make a 2d image close eyes, open mouth and so on

[main nest]

oh so nothing that big which is gonna set your PC on fire

got it 😄

[wraith jackal]

so our project is using arduino and raspberry pi with other IOT components, the project I came up with is a hologram AI voice assistant

I'm planning on using an optical illusion to make the hologram

[ornate elk]

Maybe --cuda-stream

[gentle schooner]

So what should i do to make it work?

[wraith jackal]

I don't want to create my own trained AI so I'm just gonna connect the raspberry pi to a chatgpt api

[main nest]

sounds like something I want to see how it will ended up

[wraith jackal]

I'm gonna use veadotube to create gifs with generated image from stable diffusion so I can create other animations more easily (like open mouth, blink eye, idle and so on)

[ornate elk]

Try an other workflow that uses the normal Controlnet

[wraith jackal]

https://www.youtube.com/watch?v=QPM1LwEzzwU&t=7s

this is the optical illusion im gonna use

the screen is placed on top of the box and is connected to raspberry pi

[main nest]

but tbh that is smart idea

[wraith jackal]

thanks, many people said my project idea for the class is creative

[main nest]

uni ?

[wraith jackal]

college

the name of the class is Internet Object Connected

my teacher name it abreviations IOT, idk y

Maybe internet object technology?

[main nest]

hmm I bet you will get A for creativity and A for success if everything goes according the plan, I have a good pc so if you need help hit me up

[wraith jackal]

thanks it works 🙂

[wraith jackal]

well the guy who's responsible of connecting the things together said the mic isn't good enough to receive the user voice input

[gentle schooner]

Well it's one of the only one that offer as much content for character consistency

[wraith jackal]

I just said anime girl big boobs 🙂

well it works

[main nest]

most common promp for SD users 😄

[wraith jackal]

now I can probably start generating my own thing

[wraith jackal]

🔥

[ornate elk]

You can try just to change the Controlnet model loader node

[main nest]

well Im sticking to more realistic content

[wraith jackal]

https://www.youtube.com/watch?v=yu2C3hr81NY&t=96s

this is the tutorial I'm following to create a simple vtuber

and I'll record each animations to create gifs

then I'll use event triggers in my code to activate those gifs

[ornate elk]

I'm off now gn

[main nest]

CS1o so is SwarmUI better for my goals than ComfyUI

Good night bro, thanks for the help

[ornate elk]

Swarm is comfyui Backend with an easier frontend

Both are easy to setup

[main nest]

so more user friendly version of Comfy

naah, we love the difficult part

okay bro, good night ❤️

[ornate elk]

Gn

[main nest]

are there any other supporters here

[wraith jackal]

As long as it's a SD 1.5 model I should be able to run it?

[main nest]

yea

and its not sdxl

[main nest]

i just wonder why when I do a girl it appears to have a lot of paint on her

when there is someone to answer I will be happy to wait 😄

nvm got it, I was placing in the prompt the following "(sharp:0.7), I(colorful explosion psychedelic paint colors:1.21)::0.051" and the Lora which I don't have right now "‹lora: epiCRealismHelper:1>", so I guess the fault is from this

[carmine ibex]

https://github.com/facebookresearch/xformers/actions/runs/10559887009/job/29252229610 how do I download this wheel?

[gentle sundial]

@ornate elk a question please , what's the difference between sd dynamic prompt wildcards and webui wildcards? and do they both work on all models? also how do you get the wildcard .yaml file type to work? I placed it every where and it's not working , it's making me crazy

[thorny iron]

Will the update argument in the batch file automatically up me to 3.5? Or do I need to download new version and set it up again? I use A1111.

[wraith jackal]

do you know which one should I use if I want to do an image to image?

so I can make the same model close eyes, open mouth (to speak) etc

[cunning robin]

turns out it was Hires fix the one causing the problem, i turned it off and got back working

but then how can i avoid weird faces?

[midnight roost]

hi how do i know if i have to do the step
Rename the copies to cublas64_11.dll and cusparse64_11.dll and nvrtc64_112_0.dll

[wraith orbit]

anyone here running SD via AMD on linux+rocm?

[main nest]

unfortunately no

[wraith orbit]

i have a functional install of SDforge running via rocm5.7 but id like to update to the latest 6.2 but cant get 6.2 to work at all, despite theoretically having official support for 7900XTX

[main nest]

for your system no idea tbh'

[main nest]

u have to make sure u delete the rocm5.7 before installing the new one

do you know how

sudo apt remove --purge rocm-dkms rocm-dev rocm-libs rocm-utils
sudo apt autoremove

run this command

in your terminal to remove the 5.7

sudo rm -rf /opt/rocm*

ls /opt

[main nest]

this is to delete ROCm related files

[main nest]

this to check if everything is deleted

just check if your OS is compatable with ROCm 6.2

before doing so

For example Ubuntu 20.04 or 22.04 are commonly supported

[wraith orbit]

for right now, 5.7 is only installed inside a venv folder for my current UI, which works fine.

and i believe im on 22.04

[main nest]

then all is good

[wraith orbit]

yeah im on 22.04

[main nest]

ok because its in venv you don't have to delete or remove it right now

[main nest]

can you give me your PC components like ram and video

okay so open the terminal and paste this

sudo apt-key add rocm.gpg.key

echo 'deb [arch=amd64] https://repo.radeon.com/rocm/apt/6.2/ ubuntu main' | sudo tee /etc/apt/sources.list.d/rocm.list```

[wraith orbit]

for reference this is the most common error when trying to generate via any 6.x version

[main nest]

well

[ruby yoke]

you have AMD?

[main nest]

7900XTX

if I'm correct

[wraith orbit]

^

[ruby yoke]

that would be why the HIP errror

[wraith orbit]

its supposed to have native support for 6.2

[main nest]

yea, that happens because of Incompatibility Between ROCm and PyTorch

[ruby yoke]

did you read through @ornate elk 's AMD guides yet?

[wraith orbit]

link it

[ruby yoke]

it's pinned to the top of this channel

they are pinned

[main nest]

sec

[ruby yoke]

read through all of them

[main nest]

https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

[main nest]

but the problem is with his version

if we update it to 6.2 everything should go smoothly

[wraith orbit]

im using ubuntu 22.04, SDforge webui and or any of its forks.

dropping any venv folder with 5.7 will work flawlessly

but attempting to use any version starting with 6 simply doesnt

rather i should say, by forcing any UI to use rocm 5.7, they will all work without issue

[main nest]

well you can keep the existing 5 and add the 6 to use it whenever you are in SD

[wraith orbit]

for testing and troubleshooting ive simply cloned a 2nd directory, keeping my functional ui install untouched

[main nest]

and if you want we can try to force the UI to work with 5.7 without issues

[wraith orbit]

i can make 5.7 work on any ui, thats not the issue, im trying to make any ui work with 6.2

[main nest]

but in that case there will be one little HOWEVER performance and compatibility may be suboptimal due to the newer GPU model

[wraith orbit]

my current functioning SD UI operates under 5.7

the end goal, is to be able to upgrade it to 6.2

or whatever the latest version of rocm is current

[main nest]

well then lets go

do u want to keep the 5.7

[wraith orbit]

i already have that backed up

so i can just drop it in as needed

[main nest]

its not mandatory to do so

lets just keep it for now

wget https://repo.radeon.com/rocm/rocm.gpg.key
sudo apt-key add rocm.gpg.key

echo 'deb [arch=amd64] https://repo.radeon.com/rocm/apt/6.2/ ubuntu main' | sudo tee /etc/apt/sources.list.d/rocm.list

run this in the terminal

to install the 6.2 on the main system

tag me when ready to procced

[wraith orbit]

ive done that command but it doesnt appear to have done a whole lot

[main nest]

sent me a pic

[wraith orbit]

thats the output

[main nest]

oh

that only a warning saying the apt-key is outdated

[wraith orbit]

and thats all it does.

[main nest]

thats fine

sudo apt update

do this command

[wraith orbit]

ive done that and upgrade but nothing signifcant occurs

[main nest]

ss ?

[wraith orbit]

gave me one warning

[main nest]

there is more commands to do

[main nest]

ah its ROCm repository is not properly identifying your current Ubuntu version (22.04 Jammy) and instead is expecting "Focal" (20.04)

sudo nano /etc/apt/sources.list.d/rocm.list

try this and sent me a ss

[main nest]

perfect

Change the distribution from "ubuntu" to "jammy"

tbh I hate Linux idk how you don't 😄

[wraith orbit]

who said i dont hate it?

the ONLY reason im on linux is for SD, since windows+SD+AMD is a mess

[main nest]

the only solution is this

[main nest]

save it

[midnight roost]

do anyone know why i need to do this step for zluda
Rename the copies to cublas64_11.dll and cusparse64_11.dll and nvrtc64_112_0.dll
also i have already files named that without doing it? :S

[wraith orbit]

its already saved

[main nest]

gimme one sec

[main nest]

Version Compatibility and File Path Confusion

@wraith orbit sudo apt update

lets see now

[midnight roost]

ok but if my stable diffusion already works then its fine right?

[wraith orbit]

already did update and upgrade

[main nest]

can you gimme a ss

[midnight roost]

or is it for flux perhaps?

[wraith orbit]

this was the output

[main nest]

if SD is working fine then there is no need to rename them

sudo apt install rocm-cmake rocm-device-libs rocminfo

do that, because if the system held back and didn't updated some files such as rocm-cmake, rocm-device-libs, rocminfo

this will do them manually

and then run /opt/rocm/bin/rocminfo to and sent me ss to check if ROCm is correctly installed

[main nest]

renaming the files could be necessary to match what FLUX expects

[wraith orbit]

rocminfo is returning general system information but not anything regarding a specifically installed rocm version

[main nest]

can I get the ss

[wraith orbit]

this is part of the output, again just general system info re hardware

it makes no mention of rocm

as far as i can tell, i shouldnt have a system wide install of rocm at this point

[main nest]

it looks good to me :

Marketing Name: Radeon RX 7900 XTX
Vendor Name: AMD
Compute Units: 96
Wavefront Size: 32
Fast f16 Operation: TRUE

This shows that ROCm is recognizing the GPU and is ready to use Stable Diffusion.

[wraith orbit]

which is where ive been at so far with every attempt to get rocm6.2 to function

ive done system level installs of 6.2 before and encountered the same errors

[main nest]

in that case the problem is maybe from the pytorch

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/rocm6.2

try doing that to install PyTorch with ROCm 6.2

[wraith orbit]

done that

[main nest]

and then try this ```python -c "import torch; print(torch.cuda.is_available())"

you can give me a ss after that

[main nest]

but if it returns with True, the PyTorch correctly set with the ROCm

[main nest]

have you done this before ?

[wraith orbit]

yes ive done that before too

[main nest]

btw you can double check with this ```python -c "import torch; print(torch.cuda.get_device_name(0))"

[wraith orbit]

even webui's that install 6.2 and dependencies automatically to their own venv folder fail with the same errors

[wraith orbit]

which it does.

[main nest]

can you try to run SD now and sent me a ss from the terminal

[wraith orbit]

again same error as always and as was posted above

[main nest]

did you try setting AMD Kernel Environment Variables ?

[wraith orbit]

no. not yet

[main nest]

well its worth a shot

export AMD_SERIALIZE_KERNEL=3
export TORCH_USE_HIP_DSA=1

lets see

[wraith orbit]

the top 2 vars are commonly suggested but appear nonfunctional

[main nest]

lmao i was just about to go to sleep before 1 hour but this issue seems like its not gonna allow me to sleep 😄

[wraith orbit]

it wont

because those vars also dont do a thing

that said dont lose sleep over linux's dice rolling feature where it randomly decides if it will work or not

[main nest]

can you do /opt/rocm/bin/rocm-smi and then run the UI in different terminal so we can monitor the GPU

and sent me a ss from the SD error again

[main nest]

well im not working tomorrow so it will be worth if we fix it

[wraith orbit]

the error is unchanged

[main nest]

dev's life

[main nest]

what about this

[wraith orbit]

nothing crazy

at some point im probably going to just nuke this entire install and start from scratch

[main nest]

and you say SD runs perfectly on 5.7 ??

[wraith orbit]

yes, i have an SDforge install running 5.7 in its own venv folder which operates perfectly

[main nest]

do you use SDXL

[wraith orbit]

lately ive been 99% flux

but SDXL works fine

as does any other older SD model

[main nest]

and when you get the error SD is not loading at all right

its not like you see the UI but can't do shi

im starting to think ROCm 6.2 is not fully optimized for your GPU required by models such as Flux or SDXL and maybe you will need 6.3 whenever its out

[wraith orbit]

i can load the ui

but the error occurs when i try to generate

and from documentation, rocm6.x officially supports the 7900XTX

[main nest]

what resolution are you generating in

[wraith orbit]

the error is immediate, and agnostic to generation settings

[main nest]

not the resolution, sampler or the scheduler ?

[wraith orbit]

yes, it will throw the error irregardless of generation settings

[main nest]

hmmm

have you tried with SD v1.5 or 2.1

[wraith orbit]

yes

it is safe to say that absolutely nothing inside the UI can make 6.x work

inching closer to a total OS reinstall and retry

``AMD Radeon™ Product Family Compatibility

Radeon™ Software for Linux® with ROCm 6.2.3 is compatible with the following AMD products:
Graphics Series
AMD Radeon™ RX 7900 XTX``

https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-UNIFIED-LINUX-24-10-3-ROCM-6-2-3.html

[main nest]

itching to leaving my job as dev because of linux 😄

[wraith orbit]

just for reference to show that ues, the 7900XTX has official support

[wraith jackal]

sure

@main nest

I'm looking for an art style looking like this

[main nest]

Since the error appears immediately after clicking "Generate," it points to an issue with the GPU kernel functions that are invoked as soon as the image generation process begins ROCm 6.x might not yet fully handle the RDNA 3 architecture

despite the official support claim

[wraith orbit]

you can also google and see plenty of people using this combination of GPU+rocm version

for them

they simply just installed shit and got straight to genning

[main nest]

bro

imma lose my mind over this, i tell you fr

[main nest]

so you can't run SDXL because of the PC parts however you can try SD v1.5 or v2.1

[wraith jackal]

you got model recommendation? :))

[main nest]

In these models, use the "img2img" function, which allows you to give a base image and modify it by inputting text prompts.

[wraith jackal]

yea that's what I'm doing currently

but due to the model/checkpoint I'm using not being really good, it dont give great result as the art style I'm looking for

lmao

[main nest]

ik what, i think its the best if we wait for the pro SD guy to have a nap and come back and fix it for 2 sec exactly so I can just jump off 😄

@ornate elk

[wraith jackal]

I just spent err lot of time generating nsfw pics 💀

[main nest]

hahaha its normal

[wraith jackal]

IK I MUST NOT BE THE ONLY ONE DOING THAT

[main nest]

but nsfw pics with that quality is jail time bro

[wraith jackal]

oh no no dw, I had some good quality, some bad quality

wish I had better quality

but what I notice is

when I do img2img, it's not great quality with the current checkpoint I'm using

[main nest]

hmm

[urban crown]

Trying to run txt2img.py I get this error: from imWatermark import WatermarkEncoder
I saw many people run into this from long time ago but no definite solution. Anyone runs into this? how do you fix it? I tried installing invisible-watermark but same result. Thanks in advance

[ornate elk]

Auto1111 will update itself but it can take a while to support SD3.5.
I guess forge will be faster.

[ornate elk]

Make sure you have set the hires steps to 10 and the denois to 0.5
That should work.

For face fix install the Adetailer extension

[ornate elk]

This step is optional if you want to use other webuis like comfyui

Its not bad to have them ready renamed if needed

[thorny iron]

I love you.

[ornate elk]

The wildcards extension works for everything. The dynamic prompt extension should work too

[river pumice]

im stuck on this for eternity

[cosmic geyser]

Could you please share a device requirements for sd3.5

[upper lava]

Hi Everyone!

I'm looking for some advice and possibly an answer to a question about StoryDiffusion.
Most of this stuff is new to me, so I rely on subscriptions like RunwayML and KlingAI. My interest is in img2vid and vid2vid and some lip-sync.

I recently signed up with MimicPC because they have a StoryDiffusion app. I'm not really looking to do comics or storyboards, but was very interested in the tech behind StoryDiffusion because of the promise of longer videos while maintaining consistent characters. As it stands, I'm unable to generate any video at all with in the app. Is it even able to do video yet?

Also, is the video just suppose to take the place of the still image? I'm trying to understand if it's suppose to still look like a comic strip, but with moving images.
This is a screenshot of inside the MimicPC/StoryDiffusion app. Nothing about video there, but what sold me on it was their promo page: Story Diffusion: AI Image & Video Generator for Magic Storytelling https://www.mimicpc.com/learn/story-diffusion-ai-image-video-generator-for-magic-storystelling .. Thanks in advance for any help!

[gentle sundial]

@ornate elk could anyone help me with this please , how do you get the wildcard .yaml file type to work? I placed it every where and it's not working ,

like this one https://civitai.com/models/454165/dynamic-wildcard-military-unleashed

when I extract the files I get .yaml files

I place them in the wildcards folder and they don't work

[lilac thicket]

hello

I wanna ask why my images come out like this?

[carmine ibex]

Hi, what are your settings

[idle mica]

hello everyone any one know ehy SD are forcing a landscape img saving in portrait mode?

[main nest]

what resolution are you generating in ?

[idle mica]

i fixed ahaha it was hire was setted for some reason to a expecific res idk why

[main nest]

yea

oki

[ornate elk]

What are your txt2img settings?

[ornate elk]

Wildcards have to be .txt files

[wraith orbit]

Anyone know exactly which version of ROCm I need for Ubuntu 24.04.1?

[carmine ibex]

I am trying to install xformers and am getting this:

Preparing metadata (setup.py) ... error
error: subprocess-exited-with-error

× python setup.py egg_info did not run successfully.
│ exit code: 1
╰─> [9 lines of output]
Traceback (most recent call last):
File "<string>", line 2, in <module>
File "<pip-setuptools-caller>", line 34, in <module>
File "C:\Users\z\AppData\Local\Temp\pip-req-build-p0kd4tk3\setup.py", line 548, in <module>
symlink_package(
File "C:\Users\z\AppData\Local\Temp\pip-req-build-p0kd4tk3\setup.py", line 125, in symlink_package
shutil.copytree(src=path_from, dst=path_to)
File "shutil.py", line 571, in copytree
FileNotFoundError: [WinError 3] The system cannot find the path specified: 'C:\Users\z\AppData\Local\Temp\pip-req-build-p0kd4tk3\third_party\flash-attention\flash_attn'
[end of output]

note: This error originates from a subprocess, and is likely not a problem with pip.
error: metadata-generation-failed

× Encountered error while generating package metadata.
╰─> See above for output.

note: This is an issue with the package mentioned above, not pip.
hint: See above for details.

[wraith orbit]

@main nest installed a newer version of ubuntu that supposedly has support for the newest version of rocm ive installed, and yet i am still receiving the same error as last night

[main nest]

how exactly are you trying to install it ?

[main nest]

well

in that case i think its the best to keep to 5.7

[wraith orbit]

unfortunately ive lost my 5.7 venv in the chaos of trying to set this up

[gentle sundial]

but everyone are using this .yaml file and it's working with them but not me

[gleaming mantle]

i've been avoiding SDXL and Pony, i thought i was meant to install an entirely new SD for it. Is it just getting the checkpoint and Lora? or is it more to it than that?

[wraith orbit]

you just DL the model

thats it

@main nest so ive gotten 6.2 to finally work it seems. by simply installing my usual UI to a different drive.

[carmine ibex]

I downloaded it and tried pip install

[wraith orbit]

im now copying the working install back to my preferred drive to see if it still works

and it doesnt. same error as usual.

so itll work on one drive but not another

and by drive, actually i mean partition

rebooted PC, now all seems functioning, on my end it turns out it mightve been a drive label issue? ROCm didnt like the space in the name

[tulip quiver]

Hello so i dl'ed a trained checkpoint called "flux1-dev-q8" in gguf format that i put in the /unet folder, for Forge but sadly each time when i hit the button "Generate" the pc freeze for 10 minutes & then it says "AssertionError: You do not have CLIP state dict!"

[ornate elk]

Hey you need the clip l and the t5 fp8 gguf Text encoders too

[tulip quiver]

oh okay

thanks ill try to find & downloading them

[ornate elk]

Here is the t5 fp8
https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf/tree/main

[tulip quiver]

ohw thank you

[urban crown]

For those who may run into this in the future, I got it resolved by installing invisible-watermark

[ornate elk]

Take the Q8

[tulip quiver]

this one right here need to be placed in the text encoder folder, right ?

[ornate elk]

Yes

@tulip quiver the clip l is here
https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main

[tulip quiver]

thanks mann, i appreciate your help fr

[midnight roost]

does anyone know why my microphone goes crazy when i generate a image? obv it has something to do with my gpu, is it possible to fix?

[ruby yoke]

cause you have it turned on?

[midnight roost]

so i cant prevent the mic doing it?

[ruby yoke]

if it was me, i'd just turn it off or unplug it

[midnight roost]

i mean when i talk with people on discord they can hear weird noise from my mic during the img generation so it must be something about 99% gpu util

[ornate elk]

Which settings do you use?

In txt2img

[midnight roost]

so like
Steps: 5, Sampler: Euler a, Schedule type: SGM Uniform, CFG scale: 1.5, Seed: 193169280, Size: 1216x832, Model hash: e2439cbda6, Model: satponyLightning_v10, VAE hash: 735e4c3a44, VAE: sdxl_vae.safetensors Version: v1.10.1-amd-11-gefddd05e,

[midnight roost]

or which exactly

[ornate elk]

Looks okay, did you used hires fix or not?

[midnight roost]

its not checked

and two images are 32 seconds

[ornate elk]

Okay

[midnight roost]

so i know the gpu goes ham but yea its weird the mic records it

A: 12.43 GB, R: 15.27 GB, Sys: 16.0/15.9844 GB (100.0%)

is it too much vram usage?

[ornate elk]

Maybe its the coilwine

[midnight roost]

yea i tried to use undervolt in the amd driver software so its "less" now than yesterday

[ornate elk]

You can try enable the discord noice cancel mode

[ornate elk]

Strange that its using 16gb

Did you used batch size or batch count?

[midnight roost]

batch count 2

[ornate elk]

Okay good

[midnight roost]

never learned batch size

or i dont get it

but should i try mvram in webuser?

[ornate elk]

Nope

Not needed for 16gb vram

[midnight roost]

oh ok

this is what i have in there
@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--theme dark

call webui.bat

[ornate elk]

Thats not correct

You should have --use-zluda in there

[midnight roost]

uh

then why does it work?

[ornate elk]

And --skip-ort

[midnight roost]

D:

[ornate elk]

Could be using directml instead

[wraith orbit]

im curious, are there other text encoder requirements for SD3.5?

desite having, and loading clip_l it doesnt seem to want to actually generate without throwing CLIP state dict errors

[midnight roost]

so if i change this what would happend?

[ornate elk]

You need this too:
https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/t5xxl_fp8_e4m3fn_scaled.safetensors

[ornate elk]

It shouldnt use that much vram and work faster

[midnight roost]

oh ok could u copy paste the entire thing i should replace

[wraith orbit]

i already have that one

[ornate elk]

add --use-zluda --skip-ort to the commandline_args

[wraith orbit]

with or without specifying a VAE makes no difference

cant get it to gen

[ornate elk]

You need this too
https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/clip_g.safetensors

[midnight roost]

so i replace all on that line and then add --theme dark?

[ornate elk]

Dont replace the line

Just add it behind --theme dark

[wraith orbit]

alright running with the _g and _l clips with the t5xxl_fp8, we've at least progressed to a new error.

RuntimeError: Error(s) in loading state_dict for IntegratedCLIP: size mismatch for transformer.text_model.embeddings.token_embedding.weight: copying a param with shape torch.Size([49408, 1280]) from checkpoint, the shape in current model is torch.Size([49408, 768]). size mismatch for transformer.text_model.embeddings.position_embedding.weight: copying a param with shape torch.Size([77, 1280]) from checkpoint, the shape in current model is torch.Size([77, 768]). size mismatch for transformer.text_model.encoder.layers.0.layer_norm1.weight: copying a param with shape torch.Size([1280]) from checkpoint, the shape in current model is torch.Size([768]). size mismatch for transformer.text_model.encoder.layers.0.layer_norm1.bias: copying a param with shape torch.Size([1280]) from checkpoint, the shape in current model is torch.Size([768]). size mismatch for transformer.text_model.encoder.layers.0.layer_norm2.weight: copying a param with shape torch.Size([1280]) from checkpoint, the shape in current model is torch.Size([768]). size mismatch for transformer.text_model.encoder.layers.0.layer_norm2.bias: copying a param with shape torch.Size([1280]) from checkpoint, the shape in current model is torch.Size([768]). size mismatch for transformer.text_model.encoder.layers.1.layer_norm1.weight: copying a param with shape torch.Size([1280]) from checkpoint, the shape in current model is torch.Size([768]). size mismatch for transformer.text_model.encoder.layers.1.layer_norm1.bias: copying a param with shape torch.Size([1280]) from checkpoint, the shape in current model is torch.Size([768])

its really fucking long so thats just the first bit of it

[midnight roost]

ok so like this set COMMANDLINE_ARGS=--use-zluda --skip-ort--theme dark

[ornate elk]

Oh your using forge xD
That doesn't support SD3.5 currently

[wraith orbit]

oh i see.

what UIs can use it so far?>

[ornate elk]

Make a space between them

Only Comfyui

[wraith orbit]

funky

[midnight roost]

set COMMANDLINE_ARGS= --use-zluda --skip-ort --theme dark

ok trying that then now thx

[ornate elk]

Let me know if it works

[midnight roost]

it works fine so far but i havn't tried to use flux but i can see in the cmd it says
WARNING: ZLUDA works best with SD.Next. Please consider migrating to SD.Next.

[ornate elk]

This message can be ignored

[ornate elk]

Flux is currently only supported in Forge and Comfyui

[midnight roost]

but how do i know if flux works? atm if i use flux1_dev-bnb-nf4-v2.safetensors i get pictures gray or in a mess does that mean its a wrong prompt or its not working?

[stone fog]

@ornate elk With the new SD out I believe I saw in some ad lists, is it a good time to update the AMD versioning?

[wraith orbit]

what UI and make sure your CFG is set to 1

[ornate elk]

In auto1111 you will get an error

[ornate elk]

Ad lists?
Whats your hip SDK version?

[midnight roost]

i got the cpu cuda thing error message but with that specific checkpoint it does generate its just a mess like

[ornate elk]

Yea the nf4 only works with forge webui or comfyui

[stone fog]

I’ll have to double check but should be the latest version from your AMD guide

[midnight roost]

ok so despite i can generate with nf4 it doesn't work

[wraith orbit]

if youre not on a supported UI or have incompatible prompt settings, itll give you garbage

[midnight roost]

ah ok

[cosmic furnace]

say wghat? o.o

[ornate elk]

I updated my guides a few days ago so every amd GPU should use hip SDK 6.1

[stone fog]

I just recently used GitPull so I should be updated?

[ornate elk]

Not for amd

[midnight roost]

can any give a flux prompt and settings u know that works just so i can be sure it wont work?

[atomic musk]

Anyone else just get an output that looks like this? just using the default provided workflow for SD35

[stone fog]

Is 6.1 using SD 1.5? Or am I in the wrong direction

[ornate elk]

Yep but you can also show me a CMD log if your unsure

[stone fog]

Absolutely

[ornate elk]

In Comfyui?

[atomic musk]

ye

[ornate elk]

Hmm

[atomic musk]

I did use hf transfer to grab the model 🤔 but I assume the huggingface blob in .cache is the same as the .safetensors file

[cosmic furnace]

im trying to launch forge-flux and no luck fow now =\ . that 8 gguf is starting EVERY or almost every generation for 1040 sec (yeah almost 20 mins) and generating very bad quality images

[ornate elk]

That is very strange as for me that doesn't happen

[ornate elk]

Better download the file normaly. Should be a .safetensors

Ans make sure you have all the clip text encoders too

[atomic musk]

hm, I changed some stuff

[atomic musk]

I doubt this is it. Turned out I was off of sgm uniform

[cosmic furnace]

https://civitai.com/models/630820?modelVersionId=936565 tried this one with safetensors fp8 encoder and it launched it in 1 minute and generated image with 836-1216 in 1 minute in 16 steps... bud had an error instead of image -__-

[ornate elk]

Is your forge updated? And your AMD driver updated?

[ornate elk]

1.5 is a model hip SDK 6.1 is an AMD driver

[cosmic furnace]

forge i guess ye. recently installed it... driver prior to latest.didnt installed latest one... 18 oct 24.10.1

[quartz musk]

pretty much all my images being cursed. I have it running locally but im totally new to everything .. ive changed to recommended settings that GPT told me to set .. but all my images are weird like this one

prompt - ''woman smiling slightly, laying on beach, natural sunlight, high resolution''

i cant tell if i want to marry her or pull out my flamethrower

any advice appreciated 😄

[ornate elk]

Hey, what's the current model (checkpoint) you use?

[quartz musk]

hey, sd-v1-4.ckpt

[ornate elk]

Thats the problem. This Model is 2 years old and shouldnt be used at all

[quartz musk]

ah 😛

you say Model .. the ckpt is the model?

[ornate elk]

Yep

[atomic musk]

yeah 3.5 seems to be really messed up for anatomy

[ornate elk]

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

[atomic musk]

either a huge lack of anatomy content from fine-tune or a huge limitation in the new architecture

[quartz musk]

great thank you, downloading now

delete the one I have t hen and drag the new one into my models foldeR?

[ornate elk]

Np, on civitai.com you can find many good models to try

[quartz musk]

any advice on creating images you wish you knew when you started?

all look fine?

[ornate elk]

Try less steps 35 is a good start

[quartz musk]

well thats a big improvement xD

[hallow latch]

Is it easy to update to stable 3.5?

I'm not even sure what version I am on xD

[atomic musk]

only five and a half fingers!

[ornate elk]

Looks up to date

Only torch could be a bit older

Normaly it should use 2.3.1

[hallow latch]

How can I update it from 2.1.2 to 2.3.1 easiest?

[ornate elk]

Open up a cmd and type
Pip cache purge
Then delete the venv folder

[ornate elk]

The chara searcher extension gives errors it seems

[stone fog]

Any idea what might have caused it?

[ornate elk]

Idk

Does the extension works?

[stone fog]

How exactly might I test that, stupid question probably.

[ornate elk]

Idk, you installed it xD

If you don't need it, delete it and then the extension too

[stone fog]

True

[rigid socket]

I am using SD 3.5 Large to make zoom movies. Create an image, stretch it, feed it back in as the next img2img source, repeat.
When testing SD 3.5 it seems to always lead to these canvas/grid/dots textures that ruin the look.
Probably a long shot, but has anyone else tried using SD 3.5 for these sort of movies and had any success? Maybe I am missing a setting? But all I am doing is using a slight lancosz zoom of the image between frames.
Here is an example with SD 3.5.
https://www.instagram.com/reel/DBe0MI7MZIA/?utm_source=ig_web_copy_link&igsh=MzRlODBiNWFlZA==
Here is the sort of smoother result I want to get. This one is using Playground v2 model.
https://www.instagram.com/reel/DBex6ptMgxF/?utm_source=ig_web_copy_link&igsh=MzRlODBiNWFlZA==
Here is a frame of the SD 3.5 result too as Instagram smoothed the result a bit to hide how obvious it is.
Any ideas as to why this happens or ideas to smooth out the results? The imagery "behind the canvas" looks interesting and shows some potential if it was not hidden.

Similar canvas effect with SD 3 Medium img2img, but not as severe

SDXL does not have the canvas issue and creates smooth results.

[tulip quiver]

okay so i tried but it still doesnt seem to work

i still have the Assertion Error

[ornate elk]

Can you screenshot the webui?

[ornate elk]

You have to select flux at the top left.
And then you have to enable clip l and t5
Then it will work

[carmine ibex]

I tried to install xformers but it says it can't find flash attn.

[tulip quiver]

like this ?

[stone fog]

Besides that did yous ee any issues?

Someone was telling me about I should change over to swarm UI

[ornate elk]

No

The model (checkpoint) needs to be flux

But at the encoder dropdown you can select multiple files

Choose clip l and t5

[ornate elk]

Nope looks normal. Just a bunch of warnings from extensions. But if everything works its fine.
No need to switch

[nova isle]

just joined but i hope someone can help
trying to install stable diffusion reforge and I am running into a problem

[tulip quiver]

i tried but i think the issue come from my gpu

i dont have a good enough gpu to run all these files

its a GTX 1080Ti

but thanks for the help man, ill soon upgrading my gpu in any case

[inner ginkgo]

Hope this aint a dumb question but there's nothing wrong about my commandline right? I see a hint telling me I can use --cuda-malloc but even when its in my commandline it still shows the hint at launch

set COMMANDLINE_ARGS= --share=true --listen --port --cuda-malloc --medvram --api --pin-shared-memory --cuda-malloc --cuda-stream --opt-sdp-no-mem-attention

[nova isle]

normally when this happens you just find the version of torch it's asking for but it is not telling which version of cu121 it needs

[solemn moon]

shyoh

can we reduce the clip models things' sizes

[balmy shore]

is there somewhere with a list of available command line options for sd forge?

i know ive seen it before but i just pretty much did a clean install and its a newer version so i wanna see whats avails

or are they the same as a111

oh i think they are