🤝｜tech-support

[ornate elk]

you need the qwen3 4b text encoder aswell:
https://huggingface.co/Comfy-Org/z_image_turbo/blob/main/split_files/text_encoders/qwen_3_4b.safetensors

[fast nova]

windows

[ornate elk]

ah which rocm+torch version are you on?

[fast nova]

torch 2.10.0a0+rocm7.11.0a20251128

[ornate elk]

ah okay, and image generation works, it just freezes sometimes?

forge or forge neo?

[fast nova]

Yes, it works and sometimes even give results, but ofter freezes everything so I have to reboot my PC

I don't know about forge neo, I guess it's just forge

[ornate elk]

ahh okay then the fix could be easy, in forge or forge neo just enable the Tiled VAE option at the Never oom setting in txt2img

[fast nova]

This one, right?

[ornate elk]

currently rocm+torch on windows has an bug where it timeouts the AMD driver when the VAE step is done in image generation, so tiled vae can fix that bug

[fast nova]

It didn't help, still freezed and I had to reboot my PC. But you're right, it happens after all steps of generating the image are done

[ornate elk]

do you use sdxl or illustrious models?
then try this fp16 vae file:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors

also if you use hires fix then dont leave hires steps at 0, set them to 10 or 15

[fast nova]

Yes, thanks, I'll try it

[ornate elk]

scam!

[fast nova]

I haven't tried hires with amd yet, but I usually set it to 5-8, are these values ok?

[fast nova]

It's both for IL and SDXL, right?

[ornate elk]

normal steps?

for sdxl/illsutrious 30 steps are the best

[fast nova]

Where do I put it?

[fast nova]

no, hires steps

[ornate elk]

models/vae

[ornate elk]

ah okay

10 hires steps should be the minimum

[fast nova]

otherwise I'll get AMD issues like freezing?

[ornate elk]

no just for the better quality

the freezes will come if you set it to 0

[fast nova]

I usually use 5-8 with denoising about 0.3. With more steps I see the same result and more generating time (well, on nvidia, but amd should be the same, right?).

[ornate elk]

yea it also depends if you make realism or anime etc

[fast nova]

Most of the time, I do realism.. but logically it should require more hires steps for looking good. I guess it's either becouse I use low denising or just becouse custom hires model (but I tried standard ones like ESRGAN-something and it seems to also work fine on these amount of hires steps). It also take more/same time hiresing than generation image before hires, so, I tend to minimize hires steps

Here, right? Will Forge catch it automatically or I need to set this vae as default somewhere in the settings? How will I know if Forge use this vae or default one?

[ornate elk]

you need to select it in the vae dropdown

[fast nova]

This one, right?

ok, it's now not freezing, but generating single-colored grey square

[ornate elk]

hmmm

gray or black?

[fast nova]

gray

[ornate elk]

can you try add --upcast-sampling to the webui-user.bat?

[fast nova]

Ok, i'll try

It's currently loaded something into the VRAM and doing nothing for a 2 minutes

Maybe I need to refresh browser tab and click "generate" again

It just become orange again, I'll try to restart forge and try again

[ornate elk]

hmmm okay

you could also try adding --no-half-vae

[fast nova]

Well, it again loaded something in VRAM and RAM and doing nothing (no gpu and cpu use), generate button is gray (interrupt/skip), but I don't see it's doing anything

[fast nova]

ok, I'll try that

It's the same result with both --upcast-sampling and --no-half-vae. It seems that it also tries to use gpu for a split second when I click "generate" and then cycle nothing

[ornate elk]

remove --upcast-sampling and try again

your shared memory is full, how much normal system RAM do you have?

[fast nova]

It's not full, it's kinda maximum from what was on this graph

I have 64GB of RAM

[ornate elk]

ahh okay

[fast nova]

Strange, it's stuck again with something loaded in VRAM and don't making anything in the logs

[ornate elk]

wait a bit, maybe its compiling stuff or something

[fast nova]

ok

Welll, it still not doing anything

[ornate elk]

hmm

[ornate elk]

can you try without --no-half-vae again, but this time add this into an empty line of the webui-user.bat before the @echo off
set TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1

[fast nova]

It seems to act the same. Now it's outputing progressbar before not doing anything, but that's may be just becouse I currently launched it via .bat (I'm not sure if Stability Matrix cares what I changed in webui-user.bat, so I launched it directly to be sure)

[ornate elk]

does stability matrix now has a amd rocm forge setup?

[fast nova]

Hm, interesting detail, cmd doesn't react on Ctrl+C after it not doing anything

[fast nova]

For me it seems it always has, is it isn't? You can choose which backend to use while installed.. althrough I had to delete and install torch+rocm (and also torchaudio and torch-something) manually for it to launch

[ornate elk]

yea that means they dont have it

i installed Forge Neo with rocm myself too

without stability matrix because that just adds a layer of potential issues

[fast nova]

What is difference between forge and forge neo?

Oh, I probably should ask Google about that, not you 😅

[ornate elk]

forge doesnt get much updates anymore, and forge neo is the most up to date ui based on auto1111.
It has support for more and newer models like qwen, wan, z-image, and even supports txt2video

[fast nova]

Can I move my reForge settings to Forge Neo by just copying settings file?

And does it have AMD setup? Maybe I'll try to use it instead in hope that my freezing problems will magically dissapear 😄

[ornate elk]

maybe, but it also could break stuff

[ornate elk]

i made a guide for it, but currently the rx9070 users reported that my suggested torch version didnt worked for them

so i was surprised yours worked so far

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

It would be under Forge Neo with ROCm

[fast nova]

Well, I'm not really good at python-dependencies things, but I just googled and pasted this line pip install --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all --pre torch torchaudio torchvision after removing torch , torchaudio and torchvision, on python 3.11.13. This command doesn't specify versions, so it has less chance to error something becouse dependencies version mismatch, I guess that what make it work for me. And I didn't use zluda, becouse someone adviced me to never use zluda with RX9000

[ornate elk]

python 3.13 should not be used

forge and forge neo need 3.11.9

[fast nova]

it's 3.11.13

[ornate elk]

thats okay

zluda works fine for your gpu

its more stable than rocm currently but not as fast

[fast nova]

hm, I'll try zluda and forge neo tomorrow than

[ornate elk]

forge neo doesnt work with zluda, only currently forge and auto1111

[fast nova]

ok. Didn't know it's something UI need support to

[ornate elk]

zluda emulates cuda for amd, so it has to be patched in

[fast nova]

I though it emulates cuda so well that UI don't notice is it cuda or zluda backend 😅

[ornate elk]

yea but you need custom cuda/zluda files

[heavy estuary]

i saw someone earlier mention doing it inside of the venv folder should i try there?

[ornate elk]

You can try

[marsh flare]

hello people, i need to know if i have automatic1111 or forge, how can i see what im using?

[ornate elk]

at the bottom of the webui in browser you can see the version

[copper sundial]

Its been a hot minute since i used stable diffusion. A lot has changed and I am confused. Google doesnt seem to be helping me find getting started tutorials. Where is the 'genreate' button?

[ornate elk]

which webui did you installed? xD

[copper sundial]

Ummm Vladmandic?

[ornate elk]

whats your gpu?

[copper sundial]

Im autistic and a little frazzled.Im trying to join the 'next' server but its got some weird are you human check that I have no idea

[copper sundial]

AMD

[ornate elk]

which one?

i would not recmmend using sdnext

[copper sundial]

from memory vladmandic is for the amd/radeon gpos. I used to use it fine in the past. then switched to comfy but thats naff. so tried coming back

[ornate elk]

would suggest using Forge with Zluda or Forge Neo with ROCm Guide:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

boath have the typical ui from auto1111

[copper sundial]

do you mind my asking why. It took me ages to get it up and running. Once I get started i should be fine. I managed to find the generate button. it starts makin an image but then it vanishes

Comfy has broken and stopped working for some reason. And even then it was horrible. I couldnt do inpainting/outpainting and a lot more that I used to be able to do with vladmandic

[ornate elk]

because sdnext does everything different than the auto1111 baes uis. Its settings and usage is very confusing.
Also the Forge Neo is faster and supports new models day1 like Z-image and video gen
But Guides with ROCm require a gfx11 gpu or higher

[copper sundial]

I dont understand any of that.

[ornate elk]

whats your exact gpu btw?

[copper sundial]

Umm 1 sec. 6GB AMD RADEON™ RX 6800 XT - HDMI, DP - DX® 12

[ornate elk]

Ah okay, then currently you can only used Zluda bases stuff

Rocm is still wip until it supports your gpu

[copper sundial]

Like I said(or tried to with the tism)...

I was previouslt using vladmandic which I thought was the same as auto1111.
but Im struggling mostly because the Ui has completely changed from when I last used it. So Im struggling to even get started again. When I was using it regularly, I was what some people said was 'advanced' compared to just the average person who only used simple prompts. i was able to use the pose stuff, images as a 'reference', inpainting etc. But like I said its all a new UI.

I tried using comfy but just couldnt get on with it. I found I couldnt (work out) how to do any of the stuff I used to do. then comfy stopped working so i went back to automatic/vladmandic. but after launching the ui is so different

[copper sundial]

I was under impression rocm is what I need 9im in linux)

[ornate elk]

oh your on linux ahh, yea that uses and needs rocm

[copper sundial]

I think its just the new UI im trying to learn really. like how to ooad the safety tensors(where does it save them etc) just to get it working. Like I said. I typed a basic prompt, it started generating an image(I could see it) but then it vanishes

this is an error im getting

[ornate elk]

hmm strange

maybe an vae issue

[copper sundial]

I might have launched it wrong, I might not have loaded models in properly I literally have no idea cos there seems to be no tutorials anymore

[ornate elk]

hmm, cant help much with sdnext sry :/

[copper sundial]

thanks all the same. pity I cant get into their discord

[main rune]

RuntimeError: GET was unable to find an engine to execute this computation
so close...

I should try a less demanding model first I guess

nah, this is almost definitely something I did

bruh

@pure tusk I think something's wrong

effing bot

[copper sundial]

Thanks. Ive managed to get things up and running now.

[lucid yoke]

Thats a scam bot dont join

[copper sundial]

Ah oki. I also had a message from them. I havent clicked either thankfully

[hearty crest]

How do you organize a large number of images you've generated with image generation tools? I'd love to know what tools you use for categorizing or managing them.

[heavy estuary]

aint gonna lie i was convinced first time one messaged me offering to help

[copper sundial]

Ive managed to get things up and running. Just need to know what model or samplers etc will help get better results.
Currently just poking it with a stick(changing random stuff) to see what works isnt getting me very far.

Is there some sort of guide as to what tasks(goal im trying to get to) works with what models/samples.
Being autistic doesnt help as I dont know how to phrase things

[heavy estuary]

im sitting here talking with him and then next thing i know hes brushing off my question and asking for online wallet address

[lucid yoke]

theres no set guide but if you download them at lets say civitAI theres example images showing metadata such as sampler, cfg, steps, scheduler etc

for example:"

not all images have this however

[copper sundial]

Unfortunately CivitAI is blocked where I live(UK).

So(sorry if im repeating myself). How would I go about finding how to go about what i want to do.

Goal: I want to use the left image as a guide to make images of that quality and style, but generate variants of it. As in male/female, diff clothes, race etc

[lucid yoke]

oh hmmm flux kontext or qwen edit but those are pretty beefy to run, lets say my 5080 takes 30s for a flux kontext gen

the settings are not the problem in this case, the model rather

[copper sundial]

Ahh. I think I follow.Its been a long while. So different models are better at doing specific things? (at least thats how Ive always understood it).

[lucid yoke]

yes definitely

[copper sundial]

I used to be able to load stuff in and do loras etc but like I said its been a lomg while and cognitive issues make the memory shit lol

[lucid yoke]

https://docs.bfl.ai/guides/prompting_guide_kontext_i2i heres a good example what flux kontext can do

[copper sundial]

is there a way to Dl the (safetytensor?) files from huggingface?

[lucid yoke]

if you got the hardware to run it, https://huggingface.co/Comfy-Org/flux1-kontext-dev_ComfyUI
its pretty beefy and you might wanna lookup a tutorial

im not sure if sdnext can run it

[copper sundial]

hmmm
where am I supposed to click for the file?

[lucid yoke]

files and versions

split_files

fp8 version but again make sure you can run the model before you run it

what gpu do you have?

[copper sundial]

6GB AMD RADEON™ RX 6800 XT - HDMI, DP - DX® 12

no

[lucid yoke]

oh

yeah you cant run it then

[copper sundial]

wait

16gb not 6 lol

fucked up the copy/paste lol

[lucid yoke]

definite maybe, but considering its AMD theres a chance you cant

[copper sundial]

yeah. Im using the romc thing in linux so fingers crossed

It really is a shame civitai is blocked in uk. pity they dont have the resources to to the safety stuff. that site was pretty easy to use. and had some amazing loras

[lucid yoke]

with the age verification laws also coming to europe chances are they will comply one day

[copper sundial]

found a cool lora solely for doing klingon forehead ridges and inpainted one onto myself lol

[copper sundial]

Hmm I dl it and restarted the server and its not showing up in the list of models

hang on. saved wrong folder lol

Its in the dropdown now but getting this error

annot access gated repo for url https://huggingface.co/black-forest-labs/FLUX.1-dev/resolve/3de623fc3c33e44ffbe2bad470d0f45bccf2eb21/model_index.json.
Access to model black-forest-labs/FLUX.1-dev is restricted. You must have access to it and be authenticated to access it. Please log in.

[ornate elk]

you need an HF account to download flux
but you should checkout other flux models from civitai, they are better and oesnt need and account

[main rune]

Apparently there's a known issue with ComfyUI-Zluda where I needed to disable cudnn with a node, my lack of attention got me on a wild goose chase

Everything kinda works now, really glad

[ornate elk]

nice, good to hear!

[copper sundial]

I cant use civitai they(civitai) have blocked it in UK

[ornate elk]

damn, then you need a vpn and change your DNS

[copper sundial]

Im signed into huggingface(in chrome)

[copper sundial]

cba faffing about with VPNs

[ornate elk]

what?

[copper sundial]

TLDR: CivitAI isnt going to happen.

But you did mention about having an account with huggingface

Im signed into huggingface on chrome. So am confused why the model I DL from there isnt working

[ornate elk]

it should work but you should give more information whats not working exactly

also setting up a vpn is installing a programm and clicking "ON" and then civitai works again xD
Should be easy for someone who use comfyui

[copper sundial]

1 sec. Ill get a screenshot for you(thnaks)

[copper sundial]

[sorry] something else has bugged out now. none of the checkpoints dont seem to be loading.

[cosmic furnace]

https://civitai.com/models/2168935?modelVersionId=2442439 ppl in comments seem very happy about this model. Hmgm tried it?)

" This model just destroyed all others...RIP Flux, RIP Nano Banana, RIP Midjourney, RIP Seedream 4, RIP SDLX...Crazy revolutionary tech, same as DeepSeek did in early this year 💀💀💀 Big tech CEO's are jump from their windows 🤣"

"full pants" of happines? 😄

Cs1o investigate into this. This Z image what version is better. See they say depending on GPU fp16 might be faster or nah

Or its better cuz it run on lower vram gpus? than flux-qwen?

[ornate elk]

Im using the Q8 Version, its fast and good!

9 seconds for 9 steps

[cosmic furnace]

tell me . is it any better thqan flux-qwen ?

cuz kinda... need loras to be good ?

wait 9 seconds & 9 steps? that is really good speed hm

isnt quality & accurancy awful while at it ? 🤨

[ornate elk]

It performs so well, it beats flux in performance easily.
Its from the same creators as qwen.
Prompting is very easy, tagging and normal text works.
Text generation is crazy good

The Q8_gguf uses around 17GB vram when generating

[cosmic furnace]

So its kinda upgraded qwen? 🤔 And what about this fp16 turbo version on 11 gb? Wudnt it be faster than q8 gguf ?

[ornate elk]

The Q8_gguf uses around 17GB vram when generating

[cosmic furnace]

where dyu get it ?) - arent this version better mb ? - https://civitai.com/models/2168935?modelVersionId=2442439

[whole vine]

do you also recommend zluda bases for my 16gb 7800xt? i am an absolute noob and my first try was with comfyui and im running into a ksampler error with z image

[lucid yoke]

Lets not forget with the z-image its only a turbo version so its only getting better when the base model drops

[ornate elk]

for example:
https://huggingface.co/vantagewithai/Z-Image-Turbo-GGUF/tree/main

[ornate elk]

forge with zluda will work, or forge neo with Rocm from my guides, or comfyui with rocm

[cosmic furnace]

dont tell me bf16 version uses more than 17gb vram ? o.o . And i guess the ones is on civitai site?

[ornate elk]

idk but probably

[lucid yoke]

I run the bf16 model on my 5080 and i got 10s gens, though upscaling and refining bumps it up to 30 so it barely fits in 16gb

[high talon]

hello! was wondering if any kind soul could help me, all of my prompts turn into black boxes with random colors, i know i have something very wrong, idk if its a setting or if im missing something but i cannot for the life of me find anything on google, shoot me a DM if ur goated as fuck and can help 🙏🙏 ❤️

[ornate elk]

hey, which webui, which gpu ?

[high talon]

rtx4080, umm how do i check the Web ui version, followed a youtube guide hahah

hol up

sorry i am so lost on the menus its insane

can u tell what im on from this basic ass screenshot? i jsut reset everything basicly

[ornate elk]

thats auto1111, outdated since 1 year

please install Forge Neo, guide is in the pinned messages

[high talon]

brody im boutta lose it

YESSIR

[main rune]

A year feels so long ago, I used that first to play with image gens and it was outdated in less than a month

[high talon]

its installed and im just looking for a checkpoint atm, was a painless installation, what would be a good checkpoint? im still really new to all this, flux seems recommended but idk if it makes any difference depending on what GPU i have

[high talon]

yo anybody know what the folder is called in Forge Neo where the clip for flux goes?

[ornate elk]

text_encoder

[high talon]

i ragequit it and started with dreamshaper XL instead, its very chill to learn at a reasonable pace haha, got it working

will play around and see if i can get more familiar before going back to flux, wish it was more clear where to put the 3 extra files u needed

[ornate elk]

yea if you completely new start with sdxl or based stuff on that like Illustrious for Anime

[high talon]

since ur experienced, where would u turn to find like a good fantasy art generator? that is not flux etc, like a noobfriendly one, looking to make something similar to league of legends artstyles (the very clean fantasy style)

[ornate elk]

for styles you can download and try to use Loras, they are like addons that go on top of a model to help with specific styles or characters

[high talon]

yessir, and sd1 as a tag is probs stable diff 1.5?

[ornate elk]

nope 1.5 is old

[high talon]

is there any specific reason why dreamshaper XL loves to add extra eyes and mouts etc? im using the recommended settings from civit

[ornate elk]

do you have an example?

[high talon]

this absolute HORROR (i copy pasted someone elses prompt to see if it gets similar results)

it likes adding extra limbs or fingers etc, but most uploads i see are very clean

[ornate elk]

i mean can you show the whole webui page too

with all the settings etc

[high talon]

sryy mb

[ornate elk]

np

[high talon]

i really want to learn what everything here does to be honest, like i wanna understand every slider

[ornate elk]

the issue here is that you have hires fix enabled

and by xdefault its settings are wrong

[high talon]

hmm interesting, dreamshaper xl said they had built in refiner and all u had to do was to enble hires

[ornate elk]

yea thats not true ^^

[high talon]

ah must have been mislead by this

[celest grotto]

I finally found out the problem why my stuff keeps crashing. Turned out I have been using SDXL checkpoint and my system is only 6.0 VRAM lol

Can anyone recommend me good anime checkpoint for 1.5? I'll make do with this for now until I get new PC in the near future

NSFW is a must as well.

Please

[ornate elk]

What's your GPU and which webui do you use? And how much system RAM?

[celest grotto]

I'm using a1111 webui

[ornate elk]

Ah okay, do you have --xformers added in the webui-user.bat? Thats needed for the best performance. It goes at the commandline_args line.

Also for 16GB RAM you need to increase the Windows Pagefile.
Then you can easily use the sdxl and Illustrious models

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

Also update your nvidia drivers.

I'm off for today.

[celest grotto]

I have set the pagefile. It worked for a bit before the same problem rose again.

[celest grotto]

The webui and chrome crashed on me and closed itself. The black cmd panel also closed itself.

So I didn't even have a chance to take screenshot of what happened.

XD

[fossil sentinel]

hello, i just installed forge and im getting the CUDA error: no kernel image is avalible error, i got a RTX 5080 and im pretty sure i didnt download the automatic1111

[ornate elk]

Hey, forge doesnt have native rtx50 support.
Install Forge Neo instead, it is the most up to date

Guide can be found in the pinned messages, first link

[fossil sentinel]

ahh ok! ill uninstall forge and give that a go!

tysm!

[ornate elk]

No problem 🙂

[fossil sentinel]

its this one right

@ornate elk sorry for the ping but quick question, with forge neo you launch it with webui.bat right?

[ornate elk]

Webui-user.bat

[fossil sentinel]

ahh gotcha

[fossil sentinel]

nvn i figured it out :>

[high talon]

Yo Cso1 sorry to highjack the convo, just a quick question, ima try to get flux running today again and i just had another question about the 3 extra files.

Do Text encoder5xx something and Clip go in text encoder?

And then Ae.safetnenotors goes in VAE?

Sorry for the spelling im typing at work hahaha

[ornate elk]

Yes exactly.

[whole vine]

i got forge setup now but after generating it says "failed to recognize model type". i got the z image files from this site: https://comfyanonymous.github.io/ComfyUI_examples/z_image/

is this why im getting an error?

[ornate elk]

You have to use Forge Neo, as old forge doesnt support newer models

[whole vine]

forge neo with ROCm?

[ornate elk]

What's your GPU?

[whole vine]

7800xt

[ornate elk]

Okay, yea then that

[whole vine]

i think i messed up somewhere but not sure where. i got forge neo to generate an image but its all black

[ornate elk]

You have to set it to Lumina

And then sampler, Euler simple, or multistep

[fast nova]

I installed ForgeNeo with ur guides, but have this error while launching webui-user.bat:
There's also an output of the last step before launching webui:

Hm, is it using my system Python instead one inside ../venv?

[ornate elk]

you didnt installed it with my guides because it shows stability matrix there and python 3.12, which i boath dont list in my guide

best is to uninstall python 3.12, and install python 3.11.9 as main python.
Then reinstall forge neo in a seperate folder

[fast nova]

I put it in stability matrix folder and renamed the folder for convinience, but for other things I strictly followed your guides.

[ornate elk]

yea putting anything in stability matrix folder will cause issues

[fast nova]

How do I set main python for this folder/project/webui?

[ornate elk]

only for forge neo?
edit the webui-user.bat and there you see a PYTHON=
then open up a cmd and type
where python
Then copy the path to the 3.11.9 python.exe into the wbeui-user.bat
and then you need to delete the venv folder after that before relaunching.

[celest grotto]

Hello, can you help me with the problem I said yesterday? Should I switch to use SD1.5? Or is there a way to make it work?

[ornate elk]

have you increase the pagefile ?

and added --xformers to the webui-user.bat?

[celest grotto]

As I have said above, yes, I have increased it a while a go the first time you helped ne

It worked for like a little bit before same problem rose again

[celest grotto]

Did that as well

[celest grotto]

I can’t even generate one pic

[ornate elk]

does it freeze?

[celest grotto]

Nope, it just crashed

[ornate elk]

okay, then i would suggest you try setup Forge webui, as its better for gpus with low memory.
With that it should not crash

[celest grotto]

The chrome either closed itself or turned into “aw snap”

[celest grotto]

Is Forge UI as easy to use as a1111

[ornate elk]

yea its based on auto1111

[celest grotto]

Nice

[ornate elk]

guide is on the first link of the pinned messages

[celest grotto]

Because I also read from somewhere that Forge is superior

[ornate elk]

it has its advantages yes

[celest grotto]

Okay, I’ll see to it. Thank you very much my friend

[ornate elk]

no problem, let me know if it works 🙂

[celest grotto]

So how do I uninstall the old one?

I just need to delete it righr?

[ornate elk]

yea, but save the models first so you dont neet to redownload them

[celest grotto]

Which forge should I use? Neo or normal?

[ornate elk]

normal one

[celest grotto]

I encountered an error

[ornate elk]

you didnt downloaded the forge zip or you renamed it

[celest grotto]

I unzip it into the folder

[ornate elk]

is your F drive internal or external?

[celest grotto]

external?

I think that might be the prob

[ornate elk]

yep

it needs to be on the internal

[celest grotto]

Thanks

I'll redo it

[celest grotto]

Is this normal?

[celest grotto]

Those error lines look ominous

[ornate elk]

No

Make sure you Dont run the update bat or start bat as Admin

Also make sure you placed the forge folder not on Desktop, downloads, documents, programs folder

It has to be in a newly created folder

[celest grotto]

Wait, dang, I ran the update bat. The instruction said so.

[vocal burrow]

problem is running any of them at any point as admin

[celest grotto]

Okay, I unzipped it again. Now what should I run? run.bat or webui-user.bat?

[celest grotto]

What is this "permission error"?

Oh, I managed to fix it

[fast nova]

What HIP SDK is reccomended for 9070XT on Forge Neo with ROCm? The lastest 6.4? And what drivers are recommended? I currently have 25.10.1, is it fine?

[ornate elk]

6.4 and 25.11.1 is needed

Also ROCm is currently a bit unstable on windows

You need to enable the Tiled VAE option to not get freezes

[heavy estuary]

do you know of any other possible causes for the onnxruntime error.

[fast nova]

If you mean No model named "optimum.onnxruntime" error, they has a recomendation for it in their instruction:

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

[heavy estuary]

yeah but that dont work for me for some reason

wrong reply

[ornate elk]

You can try remove --skip-ort from the webui-user.bat and delete the venv folder.
Then relaunch and apply the fix again from above

[heavy estuary]

i wanna ask just to make sure but is the zluda folder i put on my drive meant to be empty

[ornate elk]

No it should include the zluda files

[heavy estuary]

tf

[ornate elk]

Hmm looks normal

[heavy estuary]

just making sure but when you add something to a cmd you put spaces right

[fast nova]

Grok GPT said that for 9070XT (unlike RDNA3 and earlier cards) don't really work faster than DirectML. I wonder what real 9070XT users experience with performance between DirectML and zluda, because GPTs are currently really bad at knowing what "9070XT" is (and I don't blame them xd).

[heavy estuary]

just asking because i havent been putting a space while doing step 3

[ornate elk]

Grok talks BS. Because that was before hip SDK 6.4 was released xD

It added native support for rx90 cards

[ornate elk]

Yes

[fast nova]

maybe it slow becouse some of the launch parametrs and I should try to remove some of them, like --cuda-stream --attention-quad --skip-ort?

[ornate elk]

They are okay

The only thing that could broke it is stability matrix

That thing only causes issues

[fast nova]

SM doesn't do anything to webui, it only helps managing files and may cause some issues becouse of symlinks (never did I had any issues becouse of it), but it doesn't do anything with generating process at all

It's really handing for having different python version for webui than the system one, because I don't really know how to install a specific python version and manage packets specifically for it and not for the system one outside of stability matrix. And it's also good for automatically creating symlinks for checkpoints and other files that can be shared between webuis

[heavy estuary]

for me it says "module" instead of "model" is that the same?

[fast nova]

I guess it's just a typo in the instruction

[heavy estuary]

ah

i just dont understand why ive been having issues prior i installed this once before and it worked but now its just being goofy

still didnt work so im guessing somethings wrong on my end

[fast nova]

Hm, it actually don't generate anything at all if remove --cuda-stream. Without --attention-quad it shows "4.43 it/s", but actual speed is below 1 it/s

[heavy estuary]

think it did something new

ONNX failed to initialize: module 'optimum.onnxruntime.modeling_diffusion' has no attribute 'ORTPipelinePart

i got this message and now it says compilation in progress

i uh think i know what i was doing wrong

i wasnt hitting enter again after adding pip install optimum[onnxruntime], sorry for being difficult

[ornate elk]

Ohh xD I will add that to the guide lol

Compiling is good, it will take a bit

[heavy estuary]

question

is 16382 a good amount of vram

[ornate elk]

What's your GPU?

16GB is good yea

[ornate elk]

Is the webui stored on an external or internal drive?
And what about the windows pagefile settings. Is it set to system managed?

[heavy estuary]

amd radeon

rx7800 sry

[ornate elk]

Which one? xD you can check on the taskmanager under performance tab

Ah yea thats good

[fast nova]

on iternal SSD. It's surely something to do with amd adhocks, because it was fine on 3080Ti and cuda

[heavy estuary]

would i need to run the fix everytime i use sd or was it just this one time

[fast nova]

Is it set to system managed?
What does it means?

[heavy estuary]

lowkey finna paste the fix in my notepad

[ornate elk]

Only one time

[heavy estuary]

ah yeah youre right

[fast nova]

you mean pagefile's size?

[ornate elk]

Hmm okay, can you show the settings of txt2img tab while you generate the image?

Maybe its because of a to large resolution or refiner, upscaler etc

[fast nova]

There's a pagefile settings, SD is on N drive

[ornate elk]

Any reason why its limited for C?

[heavy estuary]

im confused how vram works, like do loras and models take some of it, and do i delete images to get more back

[ornate elk]

No its temporary, like ram.
Models and loras will load into it.
When the image is generated you get the vram back.

[heavy estuary]

wow

you said the first gen takes a bit right?

ight

[fast nova]

it's can't be limited. And I don't have much space for C, so I set ~6 GB as optimal maximum

[ornate elk]

Ah alright. Then yea please send the settings you use while the image generates. Screenshot of the whole txt2img page

[heavy estuary]

so im guessing its normal for the load bar to not be present first gen?

kinda takes away the suspence

[fast nova]

It should be present when it will start to generate. But before that, it should compile zluda, what takes a lot of time in the first gen. It should write something like "compiling cuda..." in the console

[ornate elk]

Can you try an other checkpoint and then with Euler a as sampler ?

[fast nova]

yes, I tried sdxl 1.0 base, but didn't tried other sampler

with euler a it's still ~1.3 it/s

[ornate elk]

Strange

Have you tried the ROCm forge neo guide?

[fast nova]

Yes, it's speed around 4-5 it/s, but it crashes the driver or reboots/freezes my pc when progressbar is full

[ornate elk]

Have you enabled the Never OOM setting ?

For Tiled VAE

[fast nova]

And here, in amdgpu forge with zluda, it also sometimes shows numbers like this, but with actual speed much below 1 it/s. I find it out when experimented with launch parametrs, but it actually happens just randomjy, disregarding launch paramtetrs. It's also freezing a lot while this happens.
Also, zluda use a lot of VRAM, more than cuda used on 3080ti. And there's OOVRAM peak in the end of the generation.

[fast nova]

Not in zluda webui. I don't remember if I checked it in rocm one

[ornate elk]

Its needed for the rocm one to not get freezes and crashes

[fast nova]

I sure I enabled it when tried rocm for the first time, but don't remember if it's still enabled on the last generations and crashes (I saved settings in default tab, but I'll try it again later with ensuring it's enabled to be sure)

[fast nova]

I tried it with "Never OOM", it still crashed my gpu driver, there's console logs and webui settings:

it also telling me that not able to install nunchaku every launch, should I do pip install nunchaku?

[ornate elk]

Nunchaku is not compatible so no need to install

[heavy estuary]

im getting a low vram warning how do i fix that

[lucid yoke]

Its a warning not an error, what gpu do you have and what are you trying to run

[lucid yoke]

Found it nvm

[heavy estuary]

im not using any loras im just trying to mess around to try and generate decent images and then use those as base settings

[lucid yoke]

Upscaling & refining could consume more vram

[heavy estuary]

all i did was crank up the sampling steps to 60 everything else is how it starts out

so am i just outta vram or how do i replenish what ive already used

[lucid yoke]

Crank the steps down to 20-30

I almost never gen with more then 30 steps

Changes are minimal

[heavy estuary]

damn bro how do you make em look good then

[lucid yoke]

Proper prompting, face fixing etc

[heavy estuary]

face fix is an add on right?

[lucid yoke]

Sure yeah for forge, adetailer or something

I use swarmui and segmenting (built in) no idea how to set it up for amd

[heavy estuary]

i could probably figure it out

thanks for the info

huh yeah really no difference the eyes just look a little whack

[lucid yoke]

Thats what adetailer is for

[shadow heron]

is there a way to use gguf models for stable diffusion zluda?

for example this https://huggingface.co/OlegSkutte/Faun-GGUF?not-for-all-audiences=true

[main rune]

use it like usual, and wait for about 45 to 60 minutes

[prisma moss]

6 is not enough

[fast nova]

I have +20 on other drive. And why do I need much anyway if I have 64GB RAM

[prisma moss]

vram or ram

[shadow heron]

hmm its not showing up as the models though even after putting it into the models folder

[main rune]

you need to load it with a gguf loader I think

and there's a bunch more stuff as well, a good tutorial would help but I don't know enough to recommend one

[tidal crypt]

Hiii! I have a question: can sd run on mobile??

Text from a reddit post i just made:

||I'm pretty new to the ai generation system, and heard that SD offers many opportunities for free if you run it on local, so i really look forward to give it a try!
The problem is: my pc is pretty bad and can't afford a new one rn, but my phone is kind of new with 16Gb RAM so i think it would somehow run easy to mid generations. Is there any way to download and run SD on my phone with all the features it has on pc?
Thank you very much!!

PS: please use simple terminology as, as i said, I'm pretty new to these things and have only surface understanding of computers||

[lucid yoke]

Short answer no, long answer even if it could it would take a long time since image generation works on vram, normal ram is painfully slow (and mobile ram is probably different aswell) maybe maybe the old 1.5 version could work but im not aware of any mobile platform

So not with all the features thats for sure but even the basic image gen is a probable no, maybe apple devices have a chance iirc

[heavy estuary]

anybody have a proper guide for installing kohya-ss?

[tidal crypt]

I actually have 16Gb of vram too, is that enough? (I don't really know how vram works)

[random turtle]

anyone got this error when running the default z-image workflow?

[nocturne sun]

Hello,
I just installed Stable Diffusion, but I’m running into an issue when generating an image: TypeError: 'NoneType' object is not iterable.
I did make sure to select the Stable Diffusion checkpoint NoobAI-XL-v1.1.safetensors [6681e8e4b1].

[ornate elk]

Hey, can you show the complete cmd log?

[ornate elk]

Can you show the workflow?

[ornate elk]

First link in the pinned messages of this channel.

[ornate elk]

You installed the webui the wrong way

Please follow the setup guide from the first link of the pinned messages here.
If you have a rtx30/40/50 GPU then install Forge Neo,
If you have an older GPU or low vram then install normal Forge

[nocturne sun]

for NVIDIA GeForce GTX 1070 is older version ?

[ornate elk]

Yea for that go with Forge

[lilac wave]

Anyone here have a DGX Spark? Is it as good as it sounds for SD?

[cosmic furnace]

How to launch z-image on neo forge? Ive got errored out after an almost of minute trying to create image

pretty sure i dont have to do this ? - Install the latest version of diffusers, use the following command:

Click here for details for why you need to install diffusers from source
pip install git+https://github.com/huggingface/diffusers

import torch
from diffusers import ZImagePipeline

1. Load the pipeline

Use bfloat16 for optimal performance on supported GPUs

pipe = ZImagePipeline.from_pretrained(
"Tongyi-MAI/Z-Image-Turbo",
torch_dtype=torch.bfloat16,
low_cpu_mem_usage=False,
)
pipe.to("cuda")

[ornate elk]

No you dont need to do that

Can you show your webui interface?

You only need 3 files or the all in one model

[cosmic furnace]

https://civitai.com/models/2168935/z-image?modelVersionId=2442439

i used this one. not gguf =\

[cosmic furnace]

downloaded all 3 files that are within this model in link

[ornate elk]

You have to set the preset to Lumina

You can also try the All in One Model from here:
https://civitai.com/models/2173571?modelVersionId=2448013
Then you wouldn't need to select the text encoder and vae.

[cosmic furnace]

did u tried lots of them?)

and what u ended up liking. if yes.

the 9gb or 11gb gguf?)

[ornate elk]

I tried the Q8+qwen3 encoder + vae and the all in one Q8 (9gb) boath work fine.

[cosmic furnace]

have u tried inpaint anything extension? Is it working? Btw ive noticed dpm2m can let u do inaccurate inpainting because it doesnt change that much or adapt to surroundings more than Euler A =\

[ornate elk]

I didnt tried inpaint anything

[cosmic furnace]

how did yu get an 9-10 sec generating result? because gguf 9 gb ?)

it generated imagee in about 15-20 sec & errored without saying why or where -_-

error at the end of generation =\

Shud i just try the version u suggested ? or whuht? 🤨

[ornate elk]

these one worked for me:

or just this one, from the link:

[still wraith]

is this discord server for StabilityAI? assuming no - but their webiste has link to this server

[cosmic furnace]

Tried BF16 version from here. Ends up just same. after 40-50 sec of trying to create image. error at the very end. withou saying any reasons why -_-

ppl seem to use it even on lower vram-power gpu's 🤨

am i missing smthing ?

The rock updated today 🤔

i had flux running no probs -___-

[cosmic furnace]

hm. i didnt used tiled vae tho. probably 😄

while using the rock 🤔

[ornate elk]

You need to enable it always for therock

Because if not it will likely crash the driver or the webui

[cosmic furnace]

RuntimeError: [enforce fail at alloc_cpu.cpp:121] data. DefaultCPUAllocator: not enough memory: you tried to allocate 9437184 bytes.
[enforce fail at alloc_cpu.cpp:121] data. DefaultCPUAllocator: not enough memory: you tried to allocate 9437184 bytes.

lol ?

and i just get an black screen o.0

2nd time generation 20 sec . nice. but... black screen again -_-

[ornate elk]

Which sampler,?

you need to use the Lumina ui preset.
Then Euler Simple, 9 steps, cfg 1

I'm off now

[tawdry ice]

Hey people
I just recently moved to Linux, and I'm currently encountering a reccuring problem. I get this error message quite frequently

||"OutOfMemoryError: CUDA out of memory. Tried to allocate 1002.00 MiB. GPU 0 has a total capacity of 15.47 GiB of which 313.50 MiB is free. Process 4829 has 134.62 MiB memory in use. Process 175903 has 6.10 MiB memory in use. Including non-PyTorch memory, this process has 14.08 GiB memory in use. Of the allocated memory 12.36 GiB is allocated by PyTorch, and 1.52 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)"||

I really wonder why, because I never had such problem when on Windows. The checkpoint/model I use is 6.9GB and my GPU is a RTX 5060 ti 16GB.
I can encounter this problem randomly, in almost any situation (during a generation batch, when generating a single image, when using ADterailer, when using ControlNet, when using none)

I use a venv with Python 3.10.6, Pytorch 2.10.0 and CUDA 12.8

[cosmic furnace]

yeah if i leave it defoult its not black anymore but ... 😄

nah when switch to euler later it either errors or black sreen. w.e the model is. lower fp8 or bf16

[tawdry ice]

It almost feels as if it's a memory leak or something. It just keeps taking more and more VRAM until it cannot do anything

[cosmic furnace]

Launched kinda, need to pick theese 2 right . Not sure which one is optimal tho

[ornate elk]

Cfg scale needs to be 1 here

[ornate elk]

Hey, on Linux you need to have a large swap file. Check that you have one or create one and set it to 32GB if not

[fast nova]

Do I need to recompile ZLUDA after driver update? How?

[ornate elk]

It will compile again after launching the webui if needed

[fast nova]

It seems to haven't been recompiled. Is it means it don't needed it? I hoped that with new driver I won't have this 1 it/s speed. Can I manually clear the compiled things?

[whole vine]

is there a setting here i can change to avoid crashes when generating higher resolutions? its able to generate 512x512 image for about 2mins but once i increase the resolution it processes until it says 10s left then my PC freezes and have to restart.

[ornate elk]

yea you have to always enable the Tiled VAE option at the Never OOM setting at the bottom of the txt2img tab

that will prevent crashes

and use the smaller Q8 gguf version or the all in one model which is 9gb

it will work much better

[whole vine]

this one? https://huggingface.co/jayn7/Z-Image-Turbo-GGUF/blob/main/z_image_turbo-Q8_0.gguf

[fast nova]

(about ROCm in Forge Neo with 9070XT). I tried it with a newest driver. Surprisingly it don't freeze my pc forever while generating, it takes few seconds for it to complete 30 steps (1024x1024, SDXL), but then it stops, adds ~3 GB of "VRAM" in RAM, all PC lags a lot about 2-3 minutes and then it completes the image. It also caused driver crash one time, but surprisingly, it still completed the image after driver crash. And after that, it crashed driver again and also crashed webui.
Also, my videocard does "pi-pi-pi" sound every time while it generating from 0 to 30 steps, I wonder what is it.

[vocal burrow]

my videocard does "pi-pi-pi" sound every time while it generating Coil whine ?

[fast nova]

Not sure it is. If it should sound like in this video: https://www.youtube.com/watch?v=HP73edpQwgc , then it's not it, it sounds kinda like old printer printing, but with a higher pitch, not as high as in the video. I never hear it in any other cases (even when generating with zluda instead of rocm). I'll try to send a file if I can attach audio in discord

[vocal burrow]

probably coil whine yes.

[fast nova]

Ok. I'm glad it not as disguasting as in the video and happens in extreme rare cases

[shadow kestrel]

Anyone know why when I'm using rgthree fastgroup muter, my groups are all squished together like this?

[civic socket]

Hi, everyone, I am working on stable diffusion + comfyUI. We are making AI animation pipeline for making animation from the storyboards.

Does anyone expert in it? If then, please contact me. Thank you

[celest delta]

So i just learned about image Z turbo... this need a text encoder... but i don't have a folder for that in my webUI... where is it supposed to go?

[ornate elk]

text_encoder

or sometimes clip

[celest delta]

yeah but i ddon't have any of those

[ornate elk]

which webui?

[celest delta]

auto1111... i did find a clip folder deep in another one, but i dunno if it's the same

[ornate elk]

auto1111 cant use z-image anyway xD

you would need Forge Neo or ComfyUI

[celest delta]

aw that's too bad

i tried to switch to another webui a while ago but it was just too different and not as intuitive as all my stuff and settings i have now

thanks anyway

[ornate elk]

have you tried Forge Neo ?

its based on forge which is based on auto1111, all have the same UI

[celest delta]

really? I'll hve to check it out then... can it install all the extensions the same way too? like tiled VAE and so on? (a life saver this one, wouldn't be able to use them otherwise)

[ornate elk]

yea you can install all the extensions, but controlnet and tiled vae are already integrated

[celest delta]

cool

[ornate elk]

@shut apex have you followed the setup guide from the pinned messages?

[shut apex]

Yes, I pretty much followed him. I just had a problem with AMD HIP SDK 6.4.2 (the version was not compatible). And what's more, I didn't understand why I had to have a ZLUDA file in addition to the SD-Zluda

[ornate elk]

then at the end you see the fix for the error you got

with optimum onnx

[shut apex]

I don't understand. What can I do to fix this ?

[ornate elk]

Step 7

[shut apex]

Ah ok. I will try this

[distant pivot]

anyone know of a node that i can use to load an mT5-small model? I have a custom built model that works, but using it with sdxl for example isn't working. making me think that a node with a different architecture is needed.

[distant pivot]

@compact fulcrum Huh? lol i didn't understand you question.

oh, i see. lol no not really. the model works, but i'm under the assumption that the architecture is so different from t5XXL that a new node is needed. mT5-small doesn't output a dim large enough for flux, from what i understand of things. so i'm currently trying to sort out how to work it in with some sdxl stuff. I figured using it with Pony models would be a bonus.

[crude halo]

anyone knows why my deforum results are so bad with SDXL and not SD15?

[cobalt agate]

That kind of looks like animatediff. That means you're using different motion models when you switch from SD 1.5 to SDXL. SDXL motion models never got out of beta.

[shut apex]

then SD works, it opens the window for me, but now when I try to generate an image I get nothing at all. Do you have an explanation?

[crude halo]

hmmm I dont think deforum works that way. It doesnt try to animate whats on the pic,, it merely generates different seeds

[ornate elk]

Can you show your webui and cmd log

[shut apex]

problem finally resolved. Thanks anyway

[tepid warren]

Died anyone know how i to add stable diffusion to openwebui?

[cosmic furnace]

So all the main problem that was on THe rock still persists it seems? only tiled vae & compiling when changing resolution 🙁

[ancient wolf]

does anyone have a good inpainting worfklow for qwen image edit (not edit 2509 model)

[lone wolf]

I have got AMD up and running, but I cant install any nodes cause of security, I've changed it to weak in the cfg but It still wont allow me to do anything in comfyui and Im very lost right now ,I cant install nodes or templates, how can I disable this permanently?

[ornate elk]

Fixing Permission Error if installing Custom Nodes or Pip packages:
Install the Comfyui-Manager, then Click on Manager > Try to install the custom node or pip package.
If you get a security warning you have to go into ComfyUI\user\default\ComfyUI-Manager and there edit the config.ini.
Change the Security Level to Weak, then save, restart ComfyUI and try install the custom node or pip package again.
After that set the Security Level back to normal and restart.

It should work like this normaly.

[frozen plover]

posting here since general chat dont allow images. I created https://sdmarks.com/ to compare inference speed of 1000+ gpus and checkpoints. It is based on Vlad's data.

[pearl narwhal]

hey guys, my 3060ti is making around 5 s/it using this model: hesperidesIllustriousV1_v10

is this an ok speed, or should it be measuring in it/s ?

[ornate elk]

wehich webui and how much RAM?

[pearl narwhal]

i have 16gb of ram, and im using auto1111

[ornate elk]

ah okay then you need to edit the webui-user.bat and add --xformers --medvram-sdxl to the line set COMMANDLINE_ARGS=
then save.

and because of the 16gb ram you need to increase your windows pagefile.
Here is the Guide how to increase it:
https://www.tomshardware.com/news/how-to-manage-virtual-memory-pagefile-windows-10,36929.html
Enable it only for C and make sure its not enabled for any other drive.
Set it to 16000 Min and 24000 Max for C
Then apply and reboot the PC.

[pearl narwhal]

is it ok for it to be more than 24000? mine is 34117

MB

[ornate elk]

yea thats okay as long as minimum is set to 16000

[pearl narwhal]

and isnt the medvram setting for 12gb? or am i mistaken?because my gpu has only 8gb

[ornate elk]

no its especially made for 8gb vram

[pearl narwhal]

oh ok

ill try it

is there a way to save my prompts so i can exit the webui and come back to the same prompt later?

[ornate elk]

nope but you can reload the last one you used when you click the lbue white arrow below the generate button

[pearl narwhal]

ok

[pearl narwhal]

its faster but its still in the s/it territory

about 3s/it now

[ornate elk]

have you restarted the PC after chaning the pagefile`?

[pearl narwhal]

i didnt since its higher than what you said, and it already was enabled only on c

[ornate elk]

and 16000mb low?

can you show your complete cmd log?

[pearl narwhal]

you mean the web ui?

can this converting from VAE to 32bit be the thing that is slowing it down? or is this expected?

[ornate elk]

remove --opt-split-attention, then save and relaunch

its expectet

[pearl narwhal]

ok

[ornate elk]

if you use this vae it wont error:
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main

also can you show your txt2img settings when generating? maybe something is wrong there

[pearl narwhal]

ok\

when i find a prompt i like i only increase the batch count

i should download this one to the VAE folder in models?

[ornate elk]

Ah the resolution is the Problem here

Illustrious best size is 832x1216

It will be much faster on that

[pearl narwhal]

ill try it, i just ran 2 prompts with 768x768 and hires fix 2x, but it didnt get faster, ill see if this res fixes it

its running 1.5s/it now

can it get faster than this?

[ornate elk]

How long does it take for a 30 steps, Euler a image?

[pearl narwhal]

lemme check

can i run the same prompt or should i do a clean slate?

[ornate elk]

Try a cleaner prompt,using more tokens also takes a bit more time

And loras too

[pearl narwhal]

i ran one with the same script, took 1min flat, ill try a prompt with just "guy in street"

it actually took longer lol

1:22

i did a second pass with just "1 boy" took 49s

[ornate elk]

Okay what you could try is to setup forge webui. It should be faster than Auto1111 for your GPU

A guide is in the pinned messages

[pearl narwhal]

the ai was really proud it even tried signing it lol

[pearl narwhal]

ill liook into it

[shadow heron]

I'm trying to follow these instructions for gguf but I don't have the text encoder folder https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1050

do I need to create a text encoder folder?

[ornate elk]

If you dont have it, yes

[shadow heron]

whenever I try to load in the flux model, it seems to refuse to select it. I assume an error occurred. where would I find that again? would it be in the command log of autoatic1111 cmd screen?

that comes along with the loadup of forge automatic1111

[ornate elk]

Auto1111 doesn't support flux

[shadow heron]

damn.....is there any other way for auto1111 to support gguf or is that not a thing

aka the web version of forge (which i i think auto1111 is)

[ornate elk]

Nope Auto1111 is outdated since over a year now

[shadow heron]

damn I need to update then....

[ornate elk]

You should switch to Forge Neo if you want the latest feature and support

[shadow heron]

does forge neo run through the same steps as auto1111 and does it support gguf

out of the box?

[ornate elk]

Yes

A setup guide can be found in the pinned messages of this channel

[shadow heron]

Nice I will take a look when I can. didn't know auto1111 was that outdated

[ornate elk]

Yea forge was the successor of it. But its also not updated anymore.
Now we have Forge Neo xD

[shadow heron]

too many editions I see

[ornate elk]

It supports the latest models like flux, chroma, wan2.2 14b, lumina, Qwen, and z-image

[shadow heron]

*never mind it is lol. I will bookmark it for later and skim through the instructions. Might not get to it for a couple of days tho, busy with some stuff.

shouldn't be too hard tho

[ornate elk]

Sure, checkout the setup guide pinned here. It makes it simple to setup

But if your on and old GPU with less than 8GB vram, its not worth.
Best is to have something like rtx30/40/50 with 12GB or more vram

[shadow heron]

I see that it only supports nvidia for neo

I have an amd 7800xt 16gb vram card

[ornate elk]

No it also supports AMD

[shadow heron]

oh ok, so ignore the nvidia label

[ornate elk]

If you checkout my setup guides its there under ROCm Forge Neo

I'm using it with my 7900xtx

I'm off now, let me know if you have any questions about it.

[shadow heron]

will let you know while skimming

[shadow heron]

since the instructions are pretty short and simple so far, I can get this done quickly @ornate elk set COMMANDLINE_ARGS, what do i put there

I only see the options for rtx

can I just use "--cuda-stream --sage --fast-fp16"

[ornate elk]

Here is the correct guide for AMD:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#amd-forge-neo-with-rocm

[trim prairie]

guys, i already install python 3.10.6 but when i run webui-user.bat, it state like this, can somebody enlighten me with this ?

[ornate elk]

Which Webui do you want to setup?

Best is to upgrade to python 3.10.11
Or uninstall 3.10.6 and then install 3.11.9

[minor latch]

best you coould go with if you dont wanna mess with stuff is using the forge https://github.com/lllyasviel/stable-diffusion-webui-forge one click package

what is the best place to post my own made extensions?

[vocal burrow]

probably #1092446741984444416

[shadow heron]

For rocm, will the option for 7900xtx work for my 7800xt. I assume it should

[ornate elk]

Yes

[mystic tide]

hi, i tried to ask something on the subreddit but my account was too new. it should be okay here, right?

i am new to this genre of ai stuff. i knew about colab and i stick with the free version for inference since it's okay for me. i got interested recently about z-image, and i can run it in there but i can't make a lora. i heard about modal on the subreddit so i'm trying it out but i can't get ai-toolkit to work in there. does anyone here have tips or a notebook?

i did try to ask llms, i used grok and gemini while making the colab notebook to import in there. but no success

the issues are that the ai-toolkit github doesn't have an updated notebook for z-image training yet, and i couldn't get grok to adapt the notebook that is there to make it work. i couldn't get the webui working either, which is what i'd have preferred. ended up wasting away credits trying alone so i hope it's an alright enough question to be asked here

[ornate elk]

Most people ask here about local installation stuff.
Can't help with collab or notebook stuff sry

[mystic tide]

oh yeah that's alright

is there any resources you could point to instead though? if not, it's fine

[ornate elk]

that was a scammer

now banned xD

[mystic tide]

lmao

[ornate elk]

but i dont have any resources on that stuff sry :/ can only help with local installations

[mystic tide]

wait, are specs stuff related? i could use some advice on what would be the ideal stuff so i can set it up on modal later

grok suggested a10g with 0.5 cpu core and 16gb ram, is that a common option for training?

[ornate elk]

a10g ?

[mystic tide]

z-image is pretty fast so the colab t4 is fine for that, but i'm really excited for the lora journey

[mystic tide]

yeah, gpu stuff. i can look up what is the equivalent for that

[ornate elk]

okay, i think z image doesnt need as much as flux or other big models

[mystic tide]

a10g is apparently comparable to rtx 4070

[ornate elk]

ah okay

yea for training the more vram the faster it is

12gb is the minimum while 16gb would be okay

cpu is not that important

ram at least 16gb

[mystic tide]

that helps a lot actually

[ornate elk]

great :), alos there are some online gpu rental service like Vast.ai or runpod.io where you can setup ai-tools or onetrainer etc

[mystic tide]

the other option grok suggested would end up costing more credits so i'm glad this will work

[mystic tide]

i've heard of those too but i can't spend money right now. that's how i saw the post about modal on reddit, $5 free credits is enough for training a lora there

i'm gonna look into those, perhaps the setup is similar and i could port it into modal

[ornate elk]

ah okay, didnt know about modal

[mystic tide]

it looks like it doesn't need to be notebook-based either, but i'm very confused about it since i'm new to this stuff anyways

[oak swallow]

How accurate is "Install Missing Custom Nodes"?
Is it possible it's returning the wrong custom node?

[shadow heron]

getting an error here for step 6

is this the cuda error?

[minor latch]

need cuda 12.x or +

activate venv

then

pip uninstall torch torchvision torchaudio
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu118

[shadow heron]

not sure why this wasn't included in the instructions lel. so I need to download cuda 12.x? but isn't htat for nvidia tho?

[minor latch]

which gpu

[shadow heron]

7800xt. was following the forge neo for amd gpus with rocm

https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#amd-forge-neo-with-rocm

[minor latch]

do you need specifically neo? neo is highly optimized for Nvidia

try looking at this

https://www.google.com/search?q=https://github.com/lshqqytiger/ZLUDA/releases

u can also use DirectML or whatever it's called but its slow

also make sure to set command line args

set COMMANDLINE_ARGS=--skip-torch-cuda-test

[shadow heron]

anything thatsupports gguf

and its neo seems to have more features and more up to date to compared auto1111

[minor latch]

comfyui supports it and its easier to install for amd

or swarmui

if you want to use it with forge you will need ZLUDA fix

[shadow heron]

so the instructions given in that link won't work or something?

that I was following with CS?

doesn't help that I never used comfyui before either

[minor latch]

im not sure i never tried running on amd

zluda is very easy

and swarmUI gives you an frontend UI aswell so no need to use comfy backend

i mean try it and you will see if it works

me personally i dont like comfy workflow aswell

altho it gives you all the control over gens u want

[shadow heron]

okay so which one do you recommend that allows me to gguf, checkpoints, and safetensors out of the box?

and easy to use and local + allows the use of my GPU

zluda sounds nice

so is forge with zluda recommended for my use then?

[minor latch]

it is your only option since you are on AMD

zluda requires some knowledge to setup but nothing hard if you want to use a111 or forge

swarmUI / ComfyUI is your best bet for out of the box. It has native oneclick installation for DirectML. But DirectML won't be as fast as ROCm but for GGUF they are actually very efficient so its okay. Try swarmUI first.

[shadow heron]

is this located int he pinned notifications of this thread

@minor latch

oh yeah it is im an idiot

[minor latch]

https://github.com/mcmonkeyprojects/SwarmUI?tab=readme-ov-file#installing-on-windows

should be oneclick install just run the BAT

[shadow heron]

i'M FOLLowing the one int the pinned notifications

is that ok

[minor latch]

never tried it, on phone rn. But i guess It doesn't get easier than one click installers

[shadow heron]

how do you disable autoloading the browser upon clicking the .bat

[minor latch]

find the launch-windows.bat or similar and --launch_mode none

for neo or a111 its COMMANDLINE_ARGS=--use-directml --no-auto-launch

[shadow heron]

I see none of that

I'll deal with this tommorow, gotta do stuff. Hopefully I don't need to migrate to a different alternative the 5th time

Still was hoping neo would work tho. Hopefully CS1 can assist further with that (will keep swarmui in my directory tho)

[edgy coral]

I get a message that says NVIDIA GeForce RTX 5090 with CUDA capability sm_120 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_50 sm_60 sm_61 sm_70 sm_75 sm_80 sm_86 sm_90

[ornate elk]

Which Webui do you tried to setup? Because that error mostly comes from outdated ones.
Try setup Forge Neo from the pinned messages.

[ornate elk]

Yes thats the cuda error. You can do the next steps then.

[minor latch]

hello, can someone check if this script does anything to their gen

i've had issues with disabling tiling but it still wont work

[dapper tundra]

i think its automatic111

is there another web ui for this that better ?

[cobalt rover]

https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides#nvidia-automatic1111-webui-stable-diffusion-webui
Hey guys, im tryna install this but when i did this yesterday on W11 i kept getting an error saying something about skipping a cuda test in regards to my gpu
Im trying again today on W10 but wanted to know if theres a specific Nvidia driver i need to update to first

[cobalt rover]

Torch is not able to use GPU, add skip torch cuda test...

[ornate elk]

Automatic1111 doesnt get update anymore. 1.10.1 is the latest version

If you have a rtx30/40/50 you should switch to Forge Neo

[ornate elk]

What's your GPU?

[cobalt rover]

Its a 3070

Damn tho i just got it workin

And youre tellin me this no longer supported?

[shadow heron]

so trying to use faun q4 gguf but this error keeps popping up. I assume Im' missing something

note i removed disable bnb from the launch parameters since the cmd wasn't recognizing that

[shadow heron]

this is the model im trying to us https://huggingface.co/OlegSkutte/Faun-GGUF?not-for-all-audiences=true

[ornate elk]

It works but doesnt get updates anymore and no new model support etc

[ornate elk]

What's faun?

[shadow heron]

a gguf image model I'm trying to use

its kinda unheard of

[ornate elk]

Can you share a link to it?

[shadow heron]

the model is the one I linked above

at huggingface

[ornate elk]

Ah okay

It seems like no webui currently supports it.
Thats why you got the error
Only a tool called dreamio, but I never heard of that.

[shadow heron]

hmm I do use dreamio and hoping to use the faun model outside of the game (which the author does state it works for kobold so I assume the same for neo forge)

still going to use neoforge for other models tho if that doesn't work

[cobalt rover]

How long has this been the case? And what are some beginner friendly alternatives?

[ornate elk]

I see its based on chroma and chroma should work on forge neo, but its based on Chroma-Flash. Maybe forge neo doesn't support that currently

[shadow heron]

hmm interesting...i assume I need a plugin

[ornate elk]

Yea flux, wan, qwen, z-image and sdxl/Illustrious work without issues

[ornate elk]

Since over a year now.
The best alternative is Forge Neo

It has the same ui as auto1111

[cobalt rover]

Okay thanks, what makes it so good?

I find a lot of this complex to get good results, so i dont mind me some guardrails

[ornate elk]

Its actively supported, and supports the latest models and has the latest performance improvements build in

Its as easy to use as Auto1111 and also supports community extensions

[cobalt rover]

Okay thanks, ill get that then

[shadow heron]

so forge noe possible won't work with faun then due to it being chroma based?

[cobalt rover]

Also if anyone is in my situation
Install chipset drivers (amd cpu) and then ddu my gpu drivers and reinstall latest
Then it worked

[ornate elk]

Yep probably. But I can test later if thats really the case

[ornate elk]

In the pinned messages is a setup guide

[cobalt rover]

Has setting up the models themselves gotten any easier in the last year?

I remember finding good configurations for them was very timeconsuming and i never really got the hang of it

[ornate elk]

Yea forge neo has presets for the model types

[ornate elk]

cant test as i dont have enough space left xD

[ornate elk]

this error seems morel ike your not using or missing the right text encoder

[quick fiber]

i got this extension

however it dont appear

https://github.com/Uminosachi/sd-webui-inpaint-anything

i am using sd-webui-forge-neo

ig it uses automatic's11111

so idk

[shadow heron]

Do you know what im missing? Tho i don't expect you to know 100 percent

Might have to ask the faun dev (creator of dreamio)

[quick fiber]

also how do u update

--------------------------------------------------------
| You are not up to date with the most recent release. |
| Consider running `git pull` to update.               |
--------------------------------------------------------

[quick fiber]

got it to work on automatic1111

however pc freezes and crashes everytime i use it

running 9070

[crude halo]

there are lots of issues with forge neo though

[crude halo]

maybe comfyui?

[ornate elk]

It works good for me and most others too

[ornate elk]

Sure, but its not as easy to use

[trim prairie]

currently installing python 3.11.9 but it say to download with microsoft store but only have python 3.12, when using it, got problem with torch...can somebody help me 😩

[minor latch]

USE ONE CLICK INSTALL ? it creates a local python enviroment so you dont need to download python at all

[ornate elk]

Don't use the Microsoft store for python. It breaks the webui.
Uninstall everything from python and install 3.11.9 64bit from the python website.
Then delete the venv folder and relaunch the webui-user.bat

[trim prairie]

thanks a lot bro

[trim prairie]

why is this ?

[ornate elk]

It would need you to have visual studio build tools installed or you to manually install insightface

But if you dont want to use faceswap stuff then you can ignore it

[celest delta]

so i just installed forge neo but as always with these tutorials it's like missing several steps and assume everyone know what they're doing
I've been told tiled VAE is auomatically installed in it, but i can't find it, and like how do you generate video? Selecting WAN at the top still only shows txt2img or img2img

[ornate elk]

Tiled VAE is at the Never OOM option

[celest delta]

i see, so it's just a check now? No option to set stuff? (i had to go down to 1024 and 64 for exemple, don't remmeber the name right now)

[celest delta]

i have soooo many question, would be nice if someone could set some time in private to help me...
Like i downloaded the new ZimageTurbo thing... isn't it supposed to generate stuff faster than other checkpoints/models? It's the same speed or even slower...

[ornate elk]

You only need to enable it yes

[ornate elk]

It needs 9 steps and Euler simple.
Ui preset lumina

[celest delta]

that's what i mean, is there a full tutorial telling you all of this? How was i suppsoed to know, lol

and it says forge neo works with video generation, but even if i click on the wan thing i don't any options for it

[ornate elk]

here is the wiki entry for each model type:
https://github.com/Haoming02/sd-webui-forge-classic/wiki/Inference-References

[celest delta]

when i follow what you said, the image are super blurry... but they were good when i tried them on wan (when i was trying to find videos, lol)

and also the speed wasn't faster

[ornate elk]

faster than what?

[celest delta]

other checkpoints/models

[ornate elk]

its definitly faster than qwen, flux and chroma and wan

but of course slower than sd 1.5, sdxl illus etc

[celest delta]

oh, i thought it was faster in general, i've been using illustrious most of the time

and how do i use video generation? i don't see the option in the WAN settings

[ornate elk]

txt2video is txt2img

[celest delta]

why doesn't it do a video then

[ornate elk]

You need at least 17 steps

And wan 2.2 14b as model

[celest delta]

but then how do i control the length of the video, amount of frames, and all that?

[ornate elk]

the wan 2.2 16b model has a fixed fps rate of 16

so 17 steps will get you 1 second of video

33steps = 2seconds and so on

[celest delta]

is there a limit?

[ornate elk]

idk

[celest delta]

also can i use other wan 2.2 models? like lower gb for lower ram? i need to find one cause the first one i found doesn't work

[ornate elk]

you can use gguf versions of wan 14b

https://huggingface.co/QuantStack/Wan2.2-T2V-A14B-GGUF/tree/main
you need to enable the Refiner in settings, then set highnoise as model and low nosie as refiner

[celest delta]

refiner is not present in forge neo, guess i have to download it as an extension?

[ornate elk]

No as I said you have to enable it in the settings

Its under refiner I guess

[celest delta]

there's a lot of different files, which one do i pick?

[celest delta]

apprently i also need a text encoder... what do i pick? ther'es not one in these files

[celest delta]

finally managed to do a generation... but it's not a video

[celest delta]

okay i didn't notice the frame slider... used that, but then the generation fails after trying to do one for 10 minutes...

i really need help, i probably downloaded the wrong wan model among all the ones in the link, and the generation slows down everything even with Lowest resolution (i have 10 gb ram), only to produce nothing at the end

[celest delta]

and does this save some stuff on your main drive? i have the webui installed on another drive than my C drive, but my C drive still lost like 20 GB from nowhere after these tries...

[ornate elk]

how much gpu Vram do you have?

[trim prairie]

guys, need help, has been stuck here for a long time now, i already install visual studio insiders 2026 but it wont do

[ornate elk]

hey, can you share the complete cmd log?

[trim prairie]

can i share it this way ?

[ornate elk]

the facefusion extensions is outdated and wont work anymore

you have to install and use the normal Standalone Facefusion version

[trim prairie]

any step how ? newbies here

one thing, any link for me to download controlnet model ?

[ornate elk]

sure here:
https://civitai.com/models/136070?modelVersionId=267493

[trim prairie]

u really are so helping, thanks a lot !

[ornate elk]

whats your gpu btw?

[trim prairie]

RTX 4050 with 6GB Vram

[trim prairie]

can i use this for SD1.5 ?

[ornate elk]

nope thse are for 1.5:
https://civitai.com/models/38784/controlnet-11-models

[ornate elk]

with that gpu you should better use Forge webui instead of auto

its more optimised for lower vram/ram amount

[trim prairie]

any idea how i can start with it ? 😅

[ornate elk]

yea, you can find the setup guide for it here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

it already has the controlnet extension included

[celest delta]

10

[lyric ingot]

Hi i just dowloaded Stable Diffusion. Ive trouble generating images. Its taking too much time idk why (10 to 15m). Ive 16go vram. Can u guys try to help me and maybe give me the best settings i should be already using ?
Thank u !

[ornate elk]

hey, whats your gpu and which wbeui have you set up?

[ornate elk]

thats not enough for wan 2.2 14b

maybe wan 2.2 5b can work, but that only works in comfyui. Or ltx video

but with 10gb vram i would stay at image gen not video

[celest delta]

only 14b work for forge neo?

maaaan...

[lyric ingot]

RTX 2070s and im using A1111 ?

[ornate elk]

ah okay, then edit the webui-user.bat
at the line set COMMANDLINE_ARGS=
you have to add --xformers --medvram-sdxl
then save and relaunch the webui-user.bat

[celest delta]

i cant wait for an easy webui to use to be able to do video... i tried swarm a while back and it was working but such a pain to use

[lyric ingot]

like that ?

[ornate elk]

no it has to be after the =

[lyric ingot]

ok

[trim prairie]

one more things, with my spec RTX 4050 6GB Vram which is suitable ? forge or forge neo ?

[ornate elk]

forge or (forge classic, which i have to the guides)

[lyric ingot]

Its working tysm.
Do u have a video to config fully Stable Diffusion i could follow ?

[ornate elk]

nope sry :/ but get yourself the Adetailer extension and boorutag autocomplete extension from the extensions tab

also go into the settings under User Interface -> Quicksettings -> and there add sd_vae
then hit apply and reload ui

then your ready

[lyric ingot]

TYSM have a great day

[ornate elk]

thx, you too!

[celest delta]

so apprently you can use wan video on forge neo with as low as 6gb ram if you use a lower Q model?

https://www.reddit.com/r/StableDiffusion/comments/1mlhh2p/running_wan_22_on_an_8gb_vram_gpu_my_benchmarks/

[ornate elk]

yes that can work

[celest delta]

but you just told be it couldn't, xD

[ornate elk]

the normal wan 14b yep. and the lower gguf versions have a bad quality and will still take ages for a 3 second video xD

[celest delta]

i still want to try something that works though

so what do i do?

[ornate elk]

you can get the lower gguf files and test. but i cant quarantee that it works. you would also need to have a lot of system ram and a large pagefile if not

or you try comfyui with the much smaller 5b model of wan2.2

[celest delta]

i have 32 gb ram

is there not a list somewhere on which model is recommended for vram? Q3, Q6? I dunno

[ornate elk]

it has to be a bit lower than your actual vram

so if you have a 3080 with 10gb i would go for a 8-9 gb version

if that errors out of vram, then go lower

[celest delta]

do you know if this works with forge neo? Since it doesn't need high or low :

https://huggingface.co/QuantStack/Wan2.2-S2V-14B-GGUF/tree/main

[ornate elk]

it should work

[celest delta]

oh wait it's S2V... what is that?

lol

[ornate elk]

its audio to video lol xD that wont work in forge

[celest delta]

oh lol, yeah

[trim prairie]

why this error appaers ? in webui it says typerror ? 😩

[ornate elk]

can you screenshot the whole webui site?

[trim prairie]

okay, i think i got it, i set the Hires steps to 0.5 😅

it should be an integral no. right ?

[ornate elk]

yep xD

[trim prairie]

😆

[celest delta]

@ornate elk

help? lol

[ornate elk]

complete cmd log pls

[celest delta]

where do i find that?

[ornate elk]

just copy the whole cmd text, there is no log

[ornate elk]

you probably dont have ffmpeg installed

its required for video stuff (converting images into video)

[celest delta]

ah okay, thanks

[ornate elk]

you need to download it and add it to the system path

[celest delta]

yeah just did that, currently generating we'll see if this works

welp... same error as before

[ornate elk]

can you open up a cmd and run
ffmpeg -version

[celest delta]

i did this tutorial to instal ffmpeg

https://www.stablediffusiontutorials.com/2025/11/forge-neo-installation.html

[celest delta]

I:>ffmpeg -version
ffmpeg version 2025-12-10-git-4f947880bd-essentials_build-www.gyan.dev Copyright (c) 2000-2025 the FFmpeg developers
built with gcc 15.2.0 (Rev8, Built by MSYS2 project)
configuration: --enable-gpl --enable-version3 --enable-static --disable-w32threads --disable-autodetect --enable-cairo --enable-fontconfig --enable-iconv --enable-gnutls --enable-libxml2 --enable-gmp --enable-bzlib --enable-lzma --enable-zlib --enable-libsrt --enable-libssh --enable-libzmq --enable-avisynth --enable-sdl2 --enable-libwebp --enable-libx264 --enable-libx265 --enable-libxvid --enable-libaom --enable-libopenjpeg --enable-libvpx --enable-mediafoundation --enable-libass --enable-libfreetype --enable-libfribidi --enable-libharfbuzz --enable-libvidstab --enable-libvmaf --enable-libzimg --enable-amf --enable-cuda-llvm --enable-cuvid --enable-dxva2 --enable-d3d11va --enable-d3d12va --enable-ffnvcodec --enable-libvpl --enable-nvdec --enable-nvenc --enable-vaapi --enable-openal --enable-libgme --enable-libopenmpt --enable-libopencore-amrwb --enable-libmp3lame --enable-libtheora --enable-libvo-amrwbenc --enable-libgsm --enable-libopencore-amrnb --enable-libopus --enable-libspeex --enable-libvorbis --enable-librubberband
libavutil 60. 20.100 / 60. 20.100
libavcodec 62. 22.101 / 62. 22.101
libavformat 62. 6.103 / 62. 6.103
libavdevice 62. 2.100 / 62. 2.100
libavfilter 11. 10.101 / 11. 10.101
libswscale 9. 3.100 / 9. 3.100
libswresample 6. 2.100 / 6. 2.100

Exiting with exit code 0

[ornate elk]

okay then idk why it has that error xD

[celest delta]

maaaaaaaaaaan there's always something wrong that nobody knows what the issue is with me... my luck sucks ass

maybe restarting the computer instead of just reloading the UI?

[ornate elk]

worth a try

[celest delta]

and as you said yesterday, if i select 33 frames, i must put 33 steps?

[ornate elk]

oh no they added the frames slider so you can leave the steps at 20 and only use the 33 frames

[celest delta]

finally it's working

took 5 minute to generate 2 seconds, could be worse!

now to find some good lora for it... do you have recommendations?

also thanks a lot

[ornate elk]

no problem, but nope no recommendation, as i dont do video stuff that much, only for testing

[celest delta]

well i just tried this one and it just give me a jumbled mess, any help for settings when using a lora with wan?

https://civitai.com/models/1872525/wan22-simple-and-pure-color-anime-style

[ornate elk]

not from me :/ maybe someone else knows

[celest delta]

i guess there's nothing i can do much about the video quality since it's a lower vram model? Can i use hi-res fix with it?

otherwise it might be better to just use image2video?

[ornate elk]

img2video might be easier
i guess hires fix wont work

[celest delta]

no change in setting for img2video? i just need to get the i2v model and that's it?

[ornate elk]

check the wiki, i didnt used img2video

[weary nacelle]

what're all the different types of stable diffusion things?
like, the comfyui, swarmui, forge, etc?

and which is probably the best to run on an amd card?

[ornate elk]

depends on your amd card

[weary nacelle]

7800xt

[ornate elk]

forge neo with rocm or forge with zluda

from the pinned guides

[weary nacelle]

oh okay, thank you

what's the difference between them all tho?

[ornate elk]

different ui and model support etc

[weary nacelle]

oh alr

[weary nacelle]

are these basically the same things?

as in are there many/any differences between the two

[vocal burrow]

It can be confusing but the core behind all of them is the same. After all, they load the same stable diffusion models (or z image, flux, whaterver).
What change is the way they load the models, the way they use your hardware, the specific knobs and settings they expose to the users, their UI.

Just like different cars can share the same engine but feel and look totally different.

[weary nacelle]

ah yea that makes sense

tysm

[vocal burrow]

Some are made to be more intuitive for users and therefore hide a lots of settings, some have a graph / node based UI, some are made with specific GPU vendors (Nvidia / AMD / Intel) in mind. Some implements specific optimisations.

It can be a mess and quite often those choices are mutually exclusives.

[ornate elk]

rocm and zluda are different backends. while zluda is more stable, it is a bit slower compared to amds new rocm support

[weary nacelle]

alr then

ty guys

[ornate elk]

no problem, feel free to ask here if any step isnt clear

[celest delta]

yeah i'm gonna need some help with I2V, the image just becomes noise immediatly

it does a video, but that's it

hope someone can help

[celest delta]

any advices on Zimageturbo loras? I downloaded one on civitai, but the generation doesn't take it into account at all, like it doesn't exist

[jade dew]

@ornate elk Today, I encountered a problem in my friend's shared session and so as in my Local SD that loras used for image generation distorts the image result. Also, when I use Forge Couple, it often occurs that the prompts prioritize the secondly added character into the image and the first one doesn't even appear on the image or she becomes almost non-existent despite her detailed prompts. Is this normal? No such thing happened to me before.

My own Colab sessions are unaffected, but what exactly is this?

[dusky osprey]

Hi people! Quick question. I'm using comfyui portable. I've a 9060XT and 32ram. I'm trying to make some clothing workflows work. Would you recommend me flux, Qwen, or any other workflow? Total noob here. Thanks in advance

[ornate elk]

Can't help much with that but it doesnt sound normal.
Check the cmd log and that all extensions are updated

[jade dew]

Forge Couple is the only extension that is not updated due to the newer version is inconvenient with certain of my characters depictions. The newest version is v5.1.0, whilst I'm only willing to use the v4.0.4 one.

[ornate elk]

what exactly did changed in v5 ?

[jade dew]

Basically, Leaves out certain elements from main prompt line and the first character lora entirely. The second character becomes the main focus somehow. In v4.0.4 it generates both characters with no problem and not willing to leave out key prompts that I added.

sometimes may cause the engine to ignore certain text-based negative prompts like english/chinese/korean texts.

[ornate elk]

you should open up an issue on the forge-couple github then, the dev answers within a day.
i cant imagine it breaks anything that much in v5