🤝｜tech-support

[placid lagoon]

I downloaded illustriousXL_v01

[wicked widget]

so thats a base model of the illustrious finetune on xl, i recommend a finetune of those, like IllustriousMiaoMiao and so on

the finetunes of base models will always be better ina llot of cases

ultimately you can use what you prefer, just sharing personal opinion

[placid lagoon]

i wanted to generate this style https://civitai.com/images/47914666

it would be nice to get closer to that

i downloaded that illustrious xl but i would need to download ntrMIXIllustriousXL_xiii.safetensor instead

right? or idk how to get closer to that result, my settings are already matching theirs

[wicked widget]

since it doesnt say which model they used its gonna be hard to tell which one

[placid lagoon]

when i copy the settings it says it here right?

Seed: 2995587984, 
Size: 896x1152, 
Model: ntrMIXIllustriousXL_xiii, 
Version: f2.0.1v1.10.1-previous-633-ge073e4ec, 
Module 1: sdxl.vaefp16, 
Model hash: c7a53a13ba, ```

[wicked widget]

oh I see didnt notice, they using this
https://civitai.com/models/926443/ntr-mix-or-illustrious-xl-or-noob-xl

[fair oxide]

how fix?

[dense tapir]

What is this error and how to fix it?

Traceback (most recent call last): File "C:\Stable Diffusion\stable-diffusion-webui\venv\lib\site-packages\gradio\routes.py", line 488, in run_predict output = await app.get_blocks().process_api( File "C:\Stable Diffusion\stable-diffusion-webui\venv\lib\site-packages\gradio\blocks.py", line 1431, in process_api result = await self.call_function( File "C:\Stable Diffusion\stable-diffusion-webui\venv\lib\site-packages\gradio\blocks.py", line 1103, in call_function prediction = await anyio.to_thread.run_sync( File "C:\Stable Diffusion\stable-diffusion-webui\venv\lib\site-packages\anyio\to_thread.py", line 33, in run_sync return await get_asynclib().run_sync_in_worker_thread( File "C:\Stable Diffusion\stable-diffusion-webui\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 877, in run_sync_in_worker_thread return await future File "C:\Stable Diffusion\stable-diffusion-webui\venv\lib\site-packages\anyio\_backends\_asyncio.py", line 807, in run result = context.run(func, *args) File "C:\Stable Diffusion\stable-diffusion-webui\venv\lib\site-packages\gradio\utils.py", line 707, in wrapper response = f(*args, **kwargs) File "C:\Stable Diffusion\stable-diffusion-webui\modules\call_queue.py", line 91, in f devices.torch_gc() File "C:\Stable Diffusion\stable-diffusion-webui\modules\devices.py", line 81, in torch_gc torch.cuda.empty_cache() File "C:\Stable Diffusion\stable-diffusion-webui\venv\lib\site-packages\torch\cuda\memory.py", line 159, in empty_cache torch._C._cuda_emptyCache() RuntimeError: CUDA error: out of memory CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1. Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

[ornate elk]

For AMD on windows follow the guides from the link in the pinned messages

[ornate elk]

More information needed. Full cmd log. When do you get the error, webui etc

[ornate elk]

Cuda out of memory.
Means not enough vram for the task you tried

[potent oyster]

tried to train a lora, getting these artifacts

was the dataset poisoned?

[potent oyster]

nvm my dumb ass just changed some parameters during generation by accident

[regal mason]

Does anyone have a good reliable inpainting workflow for SDXL in comfyui?

I've tried quite a few from various sources online and none of them seem to work properly

[river copper]

Hey guys, I'm a newbie here

How do I start creating AI art?

[midnight kiln]

Since you are here I assume you are interrested in using stable diffusion for that. There are other options too - but SD is great

you will probhably want a UI

there are a few - but I'd recommend "automatic1111"

so just google "automatic1111 installation guide"

[paper dagger]

huggingface has live model inferences too you can try. including SD3.5-Medium, Large and Large Turbo

[midnight kiln]

or you can probably find a recommended guide in here somewhere (sorry, I just joined)

[paper dagger]

there should be a sticky in here pinned for the Automatic1111 guides afaik

if people dont have high end GPUs you can try the web versions like Huggingfaces

[midnight kiln]

Ah, yes sorry - I was thinking about a local setup. You absolutely don't need that.... you can use only webssites. Its a lot easier, but there are some costs to it beyond various free usage limits

[paper dagger]

Locals the way to go but i think the learning curves a little steep for most people

[midnight kiln]

running it locally at home like me is a bit more technical - but youh ave full freedom - and it doesnt cost anything (other than you need a GPU with good VRAM)

agreed - if you haven't used SD at all before, do some free stuff on websites first

get a hang of what is possible - then consider a home setup, IF you are so inclined

I have a question of my own for more experienced users here: How can I run "highres. fix" upscaling as a separate step on an already generated low-res?

I found an extension that adds the whole thing under im2img, but then I have the problem that I can't manage to bypass the upscale&sample stage - which ruins the original picture anyway

I'm sure there is a simple solution here I am missing. I am aware I can (probably) do it as a 2-stage process via extras (for upscaling), and then do the smpling via img2img - but surely there must be an easier way. I just don't know how to use the extension properly I guess 😛

The idea is that I want to quickly gen a lot of low-res in many variations, then curate a handful of them that have potential or interresting scenes to be further refined. That seems like it would be a good way to do it if you need to do large batches - so I'm sure others have figured out how to. Share your seeecrets (plz)

[lucid yoke]

my eneregy bill went from 60 to 90 a month (234KWH to 330KWH) so theres hidden costs running ai lmao

using the power of a four person household

[midnight kiln]

I barely notice any energy use from genning - but that's probably because I have a 1080Ti. Lots of VRAm, but no tensor so it's not really able to use much of it's core. Probably running efficient in that sense... (gotta look at the positives)

I guess you can always powerlimit a card to 70% if you want to save consdiderably on costs. You'd get like 90% of the performance anyway. This is what the coinminers do, since they are actually concerned about the energy use

with a card that is actually modern and fast - you'd probably barely notice

but ye - I agree. In some cases the economical choice can actually be to use a service. Especially if you have needs like ultra-high res. where you'd otherwise need an absurdly expensive GPU like a 4090

[midnight kiln]

and a second question: How do you properly specify an OR in a prompt? I read somewhere that supposedly automatic1111 doesn't natively understand the | operator. So that would explain why I got such weird and confused gens trying it

There has to be another way though - surely?

[formal zodiac]

Uhhhhh what happened here LOL

[midnight kiln]

example: (cow|horse|man|siberian tiger|ox|man) in a field
This doesn't choose 1 of the options - instead it tries to satisfy them all - making weird horsetigers 🙂

[formal zodiac]

Is this because I generated a humanoid with a lora that wasn't trained to make humanoids? LOL

These are my settings

https://civitai.com/models/549838/phoenix-dota-2-sdxl-pony Lora here

[midnight kiln]

it looks like it finished on a noise step without refining. This can either happen with LORAs not meant for that model - or sometimes it can just happen randomly I noiced. In that case - just load another model, then back to your original one and it should clean the issue

it doesn't seem to happen unless you've been actively changing settings, so it should't do it completely randomly. I just don't know what may trigger that type of bug

[formal zodiac]

May you recommend an anime style model that would actually be compatible with this lora?

[midnight kiln]

if i can zoom the damn image in I can lol... one sec

or paste the promt for me, and say model

[formal zodiac]

Positive prompt:
masterpiece, amazing quality, best quality, lora:PhoenixDota2:0.85, female, black t-shirt, black shorts, wings on back, fire, glowing, humanoid, bird tail, tail feathers, orange skin, colored skin, headpiece, headpiece on forehead, shows face, amber eyes

Negative prompt: BadDream, badhandv4, BadNegAnatomyV1-neg, easynegative, FastNegativeV2, bad anatomy, extra people, (deformed iris, deformed pupils, mutated hands and fingers:1.4), (deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, disconnected limbs, mutation, mutated, ugly, disgusting, amputation, signature, watermark, airbrush, photoshop, plastic doll, (ugly eyes, deformed iris, deformed pupils, fused lips and teeth:1.2), text, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, masculine, obese, fat, out of frame, caricature, body horror, mutant, facebook, youtube, food, lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature

(Note: My ex-friend gave me that negative prompt to use for every image so I am not sure if the ones in the beginning are loras I don't have or not)

[midnight kiln]

Well that is a PonyXL variant - so that means any lora meant for Pony should work well. And those designed for PonyXL variants (there are several) they work even beter

[formal zodiac]

https://civitai.com/models/827184/wai-nsfw-illustrious-sdxl

This is the model

[midnight kiln]

you can sort LORA by model on civitai

ah yes, so that SDXL (stable diffusion XL). That is like the granpappy of your ponyXL variant. They can sometimes work, but they are several steps removed

[formal zodiac]

Wait

I think it's this one

https://civitai.com/models/404154/wai-ani-nsfw-ponyxl

[midnight kiln]

I'd recommend sticking to those marked for Pony

[formal zodiac]

I don't remember which one it was

[midnight kiln]

yes exactly, that's the correct variant. sometimes you are lucky and they made variants for several models

[formal zodiac]

It says pony in the model I used so

[midnight kiln]

oh - well not being able to identify the LORA is an issue

I recommend:

[formal zodiac]

I know the Lora

https://civitai.com/models/549838/phoenix-dota-2-sdxl-pony

This is the lora

[midnight kiln]

use subfolders under the LORA directory - and put the LORA into "Pony" subsolder for ony ect.

so you never mistake them or have to run around trying to look up what they are

[formal zodiac]

I just downloaded this lora today so

[midnight kiln]

and also... theres a great extension you can use in automatic1111 that scrapes civitAI for you and add metainfo to all your models and lora. Give me a sec

[formal zodiac]

Ooooh?

I wanna use ComfyUI when I make multiple characters in one image but I've noticed that the set up takes SOOOOOO long

So when I make singular characters, I use a1111

Okay I am generating my next image and I notice that the beginning looks decent at first but then as it keeps generating it gets more and more deepfried

Lemme record a video and show you

[midnight kiln]

CivitAIBrowser+

but there are a few that do a similar thing

you can grab stuff dirrectyl from automatic1111's UI - and also collect metadata on everything you already have. It makes it sooo much easier to see what the LORAs are - what the actual keywords are ect. ect. without a lot of manual sorting

and you can update everything with just a click ect.. kinda invaluable unless you literally run like 1 or 2 models only

[midnight kiln]

this is very typical for a wrong LORA

I'm 90% sure you have the wrong LORA version

just redownload the Pny one and replace the one you have if you aren't sure what you have

or use this extension I mentoned to identify everything and add metadata, images and more

[formal zodiac]

My computer froze when I tried to record but ignore that lol

Also ignore the white border

Idk how to get rid of that

[midnight kiln]

yea I see. Like I said - not uncommon

you get something close to the shape of what you wanted, but the colors (and some finer details) get fucked

just fix the LORA my boy 😄

[formal zodiac]

How do I do that?

I'm pretty sure I downloaded the latest version

[midnight kiln]

options1: reload the Pony version and replace it if your not sure you have the Pony version. you linked SDXL version first

[formal zodiac]

Yeah I linked the wrong one

[midnight kiln]

Ok, so you are sure its the right LORA version 100%?
Remove only the LORA from the promt and re-run it.
Does it still come out terrible, or does that fix it?

If yes - it's still probably a LORA issue

[formal zodiac]

What are you loading from? I don't see it when I search

[midnight kiln]

if no - it could be a more uncommon issue of FP16-FP32 coversion going wrong somewhere

[formal zodiac]

Oh nvm

I had to enable stuff

[formal zodiac]

What other extensions do you think I should install? I never installed any

[midnight kiln]

oh man og man.... there are so many good ones

im overwhelmed TBH... but I have identified some no-brainer

s

sd-webui-state-manager
For keeping your various setting between UI restars, and easily access presets

juts a massive QOL improvement

sd-webui-agent-scheduler
Lets you queue stuff as much as you want, while working on further edits... never waste GPU time 😄

[formal zodiac]

Ooooo

[midnight kiln]

sd-civitai-browser-plus
For the reasons mentioned above. Makes an organizing hell so nice and easy - and pretty with picutes no less...

adetailer
Massively helps improve details in the most needed areas, like eyes, or characters generally, or hands

once you use it - will you will never not use it again

it makes you bad gens ok, and you "good but the eyes are fucked" gens perfect 🙂

[formal zodiac]

Ooooo

[midnight kiln]

Have fun trying them out. I gotta go for a sec. you can ask me about stuff later if needed

[formal zodiac]

Okii!

[formal zodiac]

So uh, option 2?

Because option 1 didn't work

Please ping me when you get back

[ornate elk]

Thats the "Alternating Words" Feature of Auto1111:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#alternating-words

it merges the words seperated by |

[ornate elk]

you used the wrong resolution.
sdxl/pony/illustrious are trained on a 1024x1024 resolution

and your webui version is very very outdated

[formal zodiac]

Oh?

How do I update?

[ornate elk]

can you check the bottom of the webui, screenshot the versions

then i know how old it is

if its to old you need to do a clean install

[ornate elk]

@formal zodiac
if you want to do a clean install here are the Guides:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

[formal zodiac]

Hang on I’m getting dinner

[austere charm]

i upgraded my gpu from a gtx 1080 to an rtx 3060, and I'm not having performance issues but my generations aren't generating the same way they used to, even if I apply the exact same seed, prompt, settings, etc. Is there a fix or reason for this?

[supple imp]

can i get some help with inpainting, i cant get it to function. at all. at best ill get a result with no change, or i get a result with a distorted mask result. the prompt was, make the hat red.

no matter what i change, the result remains the same

[midnight kiln]

Thanks. I read though it all now - guess I should have done that from the start, but in my defense it was a lot 😛
I learned about the OR operator at least - but I really want to learn more about the details of syntax.
For example, how do you handle and AND? anime, man, blue uniform AND woman, white shirt)?
I am struggling to get some sort of consistency of grouping concepts together, and also to be precise that there are 2 subjects
Sorry if I'm ranting - you just seemed to know what you were talking about 🙂

[midnight kiln]

I am not an expert in inpainting just to be clear - I just starterd using it, but...

your problem is probably that you arent using enough denoising strength

too little and it won't make the change you want. But too much, and you'll probably get a new hat entirely 😛

and just use the prompt "red hat" I think. Not that I think that is the problem

[supple imp]

ive changed that and tried every possible setting, and that prompt aswell. the result stays the same, or just gets worse

[midnight kiln]

imm2img --> inpaint mode yes?

can you show me a screenshot of your settings used?

[supple imp]

booting sd up

[austere charm]

like if i put in a 1080 gen image into the png tab, transfer all the settings, prompt, seed, etc. over into txt-to-img, and generate it on my 3060, it just comes out slightly differently and I wanted to know if there was a way to fix that

[supple imp]

btw, the og hat was red this time, the prompt was,white hat

@midnight kiln

[midnight kiln]

check your seed settings. are they set to the same, ie. GPU?

I don't think the spesific model of card should matter - but possibly the 1080 might have some smal lvariation because it is forced to upsample FP16 to FP32 due to lacking support

some minor change probably does happen there

[supple imp]

actually i think the card might be the issue

its a amd card

rx 7900xt

[supple imp]

even used the basic model

[midnight kiln]

something along those lines?

or a purely color change?

I had to use loopback to get this to work, otherwise it wouldnt change the color over entirelly, just add red details

not sure why something this simple was playing hard to get 🙂

[supple imp]

the dude is just a test lol i generated that to better show the problem

i found someone with the exact same issue

it might be an amd thing

it does the change it needs to during the live part, but at the end it just stops

[midnight kiln]

there is probably a smarter way to do it than I just did

and excuse the poor quality - that's not SDs fault, just ran it super quick

I have a shitty 1080 Ti 😦

[supple imp]

WTF

OMG

OMG

i switched to chrome

and it worked

i have been troubleshooting for the past week

[midnight kiln]

I have a hard time seeing why that would matter

what did you use before?

[supple imp]

firefox

[final vector]

Would I be able to run FLUX .1 S/D models on Arc A770 16GB?

Also is there a large difference between S and D in terms of quality?

[midnight kiln]

I'm not familiar with those versions. havent used flux much

but the page probably describes it

16GB is more than enough for most things

it mostly depends on what sort of resolution you want to work with - less so the model

With my 11GB I can't really go above 1024x1024 upscaling unless I use some sort of tiling process - but that's not ideal

[wicked widget]

yes you can

use wavespeed for additional speed

on comfyui

flux outside of comfyui is NOT worth it at all.

[midnight kiln]

how is comfy compared to automatic? I've been learning new things like mad - so ive not had a chance to compare 😛

[final vector]

😭 Time to learn how that works

I feel like Comfy is the go-to for those who want to actually automate their image creation

[wicked widget]

Comfyui is what the more advanced users use nowadays, it offers better performance under the hood and complex workflows

[final vector]

Makes sense, I just need to learn how to use it.

[wicked widget]

I have some workflows for you all to start with.

[final vector]

Is it fairly complicated?

[wicked widget]

save these .jsons and drag and drop them in comfy for it to load

it seems complicated at first

but its not imo

[final vector]

I've use the Graph Systems in Blender, but I hve no idea how it compares.

[wicked widget]

Comfy Manage addon is almost needed to make sure custom worfklows from others work as it will detect missing nodes and missing fils

when comfy is installed, go to the custom_nodes folder and then git clone https://github.com/ltdrdata/ComfyUI-Manager.git

restart comfy completely

[dense tapir]

Thank you for the response

[burnt salmon]

is forge compatible with TensorRT?

[wicked widget]

no

use a1111 normal if you want to use tensorrt

[burnt salmon]

ok thanks

[final vector]

I will set this up tomorrow, are you okay if I ping you for help at any point?

[wicked widget]

yup

if i dont reply im sleeping and ill eventually get to you

[final vector]

kk no worries, thank you!

[wicked widget]

my pleasure

[final vector]

Oh yeah I meant to ask, Arc A770 16GB, can that run Flux-Dev?

[wicked widget]

yup

i run flux dev on my 3060 12gb

but i also use wavespeed for first cache block to help w the speed

with it and xformers i can get speeds up to 1.62s/it

without it, 3s+/it

[midnight kiln]

Hey, can someone tell me how you can skip past the latent upscaler step in img2img? it craps all over gens that are already upscaled. Adetailer ahs an option to spesifically skip the step - but what about everything else?

[wicked widget]

how high is the denoise

higher denoise = more changes, reimagining the image

lower denoise = less changes, more improving whats already there

[neat slate]

hi

[manic jewel]

"When I try Google corabo, install the extension, and start it again, it says ""An error occurred while loading the script"" in the code."

If you know how to solve this problem, please give me some advice

[wicked widget]

im pretty srue google colabs stopped working for sd awhile ago

[manic jewel]

返答ありがとうございます

[wicked widget]

you're welcome

[cerulean kite]

Hello guys，when I encountered some difficulties when starting stable diffusion in my stability matrix for mac as follows: when local host is not accessible, a shareable link must be created. Please set share=True or check your proxy settings to allow access to localhost.How can i solve this problems ?

[wicked widget]

its already running on local

look at line 24

and click it

if you done so and still giving error

click the cog wheel

then enable share

[cerulean kite]

Very grateful！Very accurately solved my problem.❤️

[wicked widget]

my pleasure

[cerulean kite]

i get another stupid problem,like this : Warning: caught exception 'Torch not compiled with CUDA enabled', memory monitor disabled,honestly i've used windows for almost my life,i really feel powrless to macos

[wicked widget]

delete the ui entirely, and when reinstalling hit the advanced tab and ensure this is enabled

[cerulean kite]

I'm so stupid! Thanks to you🙏

[wicked widget]

np

[ornate elk]

There is no way to fix it. Its because you have a different GPU now. That changes the outputs sightly. This can also be caused by using xformers or a VAE as they make very little changes in the output

[ornate elk]

Make sure you whitelist the webui in any browser adblocker and it should work in Firefox too then

[ornate elk]

Do you use DirectML or Zluda?

[ornate elk]

Its a bit difficult to get 2 different characters onto the image and not merging the details.
For auto1111 there is an extension called Regional prompter. With that you can do that much better.

[manic jewel]

Please help me

Extension not available due to above error

[umbral relic]

Hello, used auto1111 10hrs ago and it was working fine... now i get a torch error: RuntimeError: Torch is not able to use GPU. Now I am trying to reintall the venv (deleted the venv folder) but still... why it broke suddently?

Just started 2 games to be sure my gpu (rtx2080) is working fine.

What happened?

Wt?

[ornate elk]

can you show your extension folder content?

mostly an extension broke the webui

[umbral relic]

But was working yesterday?

Added riffusion today but even if I remove it... its not working

Trying a fresh install from github

[vocal burrow]

deleting extension folder is often not enough as they install stuff to your venv

when deleting extension I recommend deleting your entire venv too. (it will reconstruct itself at launch)

[umbral relic]

Yes, made this step but gave me still the same error. Just to point out: I bricked my 3080TI less than a month ago... you understand that i was a little scared but a second installation (webui forge+flux) was still working so... I think auto1111 just broke by itself. Like I generated images, got to sleep (with the pc running) and this morning it was not working anymore (just close it and started it again).

Pc restart made no changes.

[vocal burrow]

not sure what you mean by "bricked my 3080ti". Especially if secon installation of forge worked.

[umbral relic]

It's a different story, bricked my 3080 gaming on rust, nothing to do with SD

[vocal burrow]

Sure stable diffusion is taxing on your gpu. But it's made to handle that kind of charge. Just got to make sure the temps are correct (just as with gaming)

[umbral relic]

But still I was scared today that maybe I briked my 2080 too

[vocal burrow]

oof. you cooked it ?

might be time to add some fan to your case :/ and/or check how to undervolt (safely) your gpu.

[umbral relic]

But still:

[vocal burrow]

full log ?

[umbral relic]

where is the full log?

[vocal burrow]

in the very console you're screenshoting

there's no file loging

you just have to copy paste the whole thing

[vocal burrow]

cf the very start of the log

SD is not compatible with python 3.11

must be 3.10.6 or above (3.10.7, 3.10.8, etc)

[umbral relic]

Why it was working yesterday? I had this error for ages now... I just ignored it but everything was working

[vocal burrow]

I doubt it was happily working

[umbral relic]

maybe you are right but... things broke this night. Everything was working 12 hours ago. No changes.

Ok I try to downgrade python

[vocal burrow]

either :
1/ pure luck, because yes it can work with 3.11 but it's really not recommended, not tested for it, it will work until the moment it call some function that has changed from 3.10 to 3.11, etc
2/ you/something installed python 3.11 recently and made its prio above your 3.10 install

[umbral relic]

You think something updated automatically this night and broke everything? is the only thin I can think

[vocal burrow]

maybe, unlikely but maybe

[umbral relic]

I try and give you fb if working or not

[vocal burrow]

unless you installed python through microsoft store (please don't)

[umbral relic]

No 🤣

[vocal burrow]

I wouldn't be surprised if it updated stuff behind your back then

once you've downgraded your python you'll have to delete your venv folder (as it's been built with your previous python)

[umbral relic]

I just hope the python downgrade will not break other python stuff on my pc... will be auto1111 one day updated to a recent python version?

[vocal burrow]

maybe, don't know. If updated python is necessary for something else you can keep both install at the same time you just have to point A1111 / forge / etc to the correct 3.10 one in the webui-user.bat

[umbral relic]

The problem was riffusion, don't install it!

Like you said... it breaks something and even uninstalling it is not helping at all

[wicked widget]

Riffusion still broken? that thing been broken for so long now

[umbral relic]

Yeah... any alternative?

Wanted to make some music for the game I am developing... but can't find any good free alternative.

[wicked widget]

stable audio

[umbral relic]

It's an extension or frontend that I can install locally?

Found a stable audio 2 demo github but idk if this is what you mean

[wicked widget]

https://comfyanonymous.github.io/ComfyUI_examples/audio/

[umbral relic]

never used confyUI... maybe time to check it out.

[junior geyser]

Anyone here use Zluda? Mine keeps saying this

[wicked widget]

as it says

its compiling

give it time

[half island]

Hoi, i'm mixing a friend's face, but if i use denoise less than 70, which gives way too different results, doesn't blend well at all, all the inpainted area gets a way too foggy result. Why is that?

Using latest webui forge.

[half island]

Settings i use

[wicked widget]

100 steps? extremely high cfg, high denoise, everything is just..cranked up

[lucid yoke]

^ try lowering all those

[formal trout]

I have a problem running pulid in Comfy

an unsupported architecture! Overriding blas backend to hipblas (Triggered internally at /pytorch/aten/src/ATen/Context.cpp:328.) return torch.nn.functional.linear(input, weight, bias) !!! Exception during processing !!! Traceback (most recent call last): File "/home/mario/ComfyUI/execution.py", line 327, in execute output_data, output_ui, has_subgraph = get_output_data(obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb) File "/home/mario/ComfyUI/execution.py", line 202, in get_output_data return_values = _map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb) File "/home/mario/ComfyUI/execution.py", line 174, in _map_node_over_list process_inputs(input_dict, i) File "/home/mario/ComfyUI/execution.py", line 163, in process_inputs results.append(getattr(obj, func)(**inputs)) File "/home/mario/ComfyUI/custom_nodes/pulid_comfyui/pulid.py", line 240, in load_insightface model = FaceAnalysis(name="antelopev2", root=INSIGHTFACE_DIR, providers=[provider + 'ExecutionProvider',]) # alternative to buffalo_l File "/home/mario/.local/lib/python3.10/site-packages/insightface/app/face_analysis.py", line 43, in __init__ assert 'detection' in self.models AssertionError
No idea, I tried to follow the installion of insightface, but as Comfy don't run with a venv, I pip installed in the directory

I tried to follow this

[hot thorn]

Hi. I have a problem with going a step further with details and quality in my pictures.
I pretty much finally completed a perfect prompt for majority of the stuff I want to do. It produces great results at good quality, but I want to make details even better, however when I change my initial settings, 30 steps, 10 hires steps, latent antialised, 1080x1080, upscaled 1.5, denoising 0.3 to something like upscaled x2, or 1200 and upscaled 1.6-1.8 it messed things up, greatly deforming my character, mainly messing up torso, or head, adding extra limbs etc.
What can I do to avoid this deformities? Changing denoising to 0.5 does nothing. Is it maybe the fault of the upscaler I use? So far I only used Latent (antialiased), since it produced best results for me over the weeks.

[hallow latch]

I am wondering, when I do a prompt.. maybe like 10 images. Is there a way for only those prompt to end up in their own folder that I decide .. Like when you do upscaling and such? Like the thing with VAE that you can change quick and simple?

[civic socket]

Wondering if someone can help me with a prompt. can't get an image of a full-body of this character for nothin. Keeps croppin him

[supple imp]

DirectML

W support btw, actually scrolled up to help people youre a goat

[formal trout]

It is normal to take 7 minutes to make one generation with Flux with my RX 6700 on Ubuntu?

Also that message

[hallow latch]

I am trying to do Marge Simpsons but the ADetailer is not working? :/

[ornate elk]

Looks good

[ornate elk]

Would strongly recommend following my Zluda guide.
Your getting a much better performance and are no longer limited on resolution or model sizes

[gleaming niche]

Hi, i am trying to use webui and stable diffusion but it tells me i need an nvidia driver. the problem is my gpu is an Intel(R) Core(TM) i3-N305, 1800 Mhz, 8 Core(s), 8 Logical Processor(s) and my display adapter driver is Inter(R) UHD Graphics. i am using a laptop. it is out of the cards for me to get an nvidia graphics card and i fear trying to install an nvidia driver will fry my laptop. i could just use another stable diffusion program thats online but i want to use webui. what is the solution?

i know its probably gonna be really slow but this laptops my only option to use webui

[karmic vortex]

Not an expert, but I don't think it would wotk on a Intel UHD graphic...

[gleaming niche]

hmm

[karmic vortex]

It says only Nvidia or AMD as GPUs

[hallow latch]

He is another bot @formal trout

[gleaming niche]

my best option is probably to ditch webui and use something online

[hallow latch]

Do not accept and join that discord!

[gleaming niche]

oh well

[hallow latch]

They will make you pay for help.

[gleaming niche]

i might as well ask how to use a cloud gpu

[gleaming niche]

@karmic vortex ?

doesnt have to be webui

i meant like just link to one program

[karmic vortex]

Not sure, not an expert, I'm kinda new here. On stable difussion webui I see a link to an external link... maybe there's a way

[brazen sedge]

I keep getting this error message , What can i do to fix this?

[gleaming niche]

this guy doesnt have an nvidia driver either

well it was worth a shot

[karmic vortex]

Check their readme.md. There's a link for Intel GPUs... but they do not provide an official support 😦

It says "External wiki page"

Maybe there's hope for you 🙂

[gleaming niche]

https://github.com/openvinotoolkit/stable-diffusion-webui/wiki/Installation-on-Intel-Silicon

[lone wolf]

Any tips on what UI to use and how to use flux? Bought myself a used 4090 and thought I should get it running, but it broke the current SDXL I use, so probably need to replace it all :S

Big diff with Forge / ComfyUI?

[brazen sedge]

I believe it is getting worse but I'm unsure what I've done to cause this blunder.

[ornate elk]

What's your GPU?

[ornate elk]

Rent a GPU online or use services like Leonardo ai

[signal jasper]

Hi all, I am using the "Save Image" node with the prefix "gym/image_%KSampler.seed%" - and the tutorial I followed shows that this saves the filename as image_SEED# however its simply saving the filename as image_%KSampler.seed%00002.png - Does anybody know what I am doing wrong? Thx

[supple imp]

Can you slide me a link?

[hot thorn]

@ornate elk is there a place where I can download upscalers? I like latent antialiased, mainly because of the slight blur which makes my stuff look very nice, but it doesnt allow me to go beyond 1080x1080 upscaled by 1.5, since at that point it deforms bodies and limbs quite a lot. I tried some 4k upscalers which work fine even when i go to 2160x2160 (after upscaling x2), but theyre way too clean and i dont like it much. is there some latent upscaler that goes to higher resolutions without deformities? Or is there something I can do to make my current upscaler work with higher resolutions? My current setup for generation is: Stable Diffusion Reforge, 1080x1080 resolution, upscaled by 1.5, 30 steps, 10 hires steps, CFG 5, denoising strength 0.3, Using Euler A with automatic schedule type.

[normal wedge]

hello so i am quite new to stable diffusion. i am using the webui stable diffusion. i was able to get it working and generate image just fine but it is taking a very long time to generate the image. There was some times it could run fast like 1-2 minutes per image for 500 x 500 30 sampling steps. but a small tweak in the prompt could suddenly make the image generation like 5-15 minutes i am not sure why. I am running this on a laptop with GPU RTX 4060 32GB of ram so idk what is the cause.

[neat slate]

你

ni men hao aiyou

[ornate elk]

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

Go for any of the Zluda versions

[ornate elk]

Here is the upscale model site, but they dont have latent upscalers.
https://openmodeldb.info/

You can use the latent upscaler in img2img with the SD upscale script to get larger resolutions

[ornate elk]

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

[coral ermine]

Hello.

I have a service (t.me/neurotools_bot) providing access to different SD models. But there was a bug in my code. I used HTTP header stability-client-id instead of stability-client-user-id. I.e. I thought that stability-client-id is for user ids.

Because of PDF-files my account is now blocked. I wrote a e-mail letter to support, but they are silent. (may it be because it's Sunday?)

Now I've fixed that bug and now user ids are sent properly. Maybe here will be an answer.

P.S. It's all (my acc. is blocked) because my program code has been sending user ids wrong.

[lucid yoke]

this is a community server. yopud be better off to email stability it self.

[coral ermine]

I did yesterday morning, but no answer.

[lucid yoke]

yes useuallty businesses have a respond time of one work week

and it was a saturday

so id expect a reply around thursday/friday

[wicked widget]

but yeah as this is a community server the ppl youre most likely looking for wont be here to see

we help those running SD locally

[lucid yoke]

and seeing how it was used by PDF's you might have bigger issues if they didnt contact law enforcement tbh

[wicked widget]

👀 now im intrigued

oh i just got the slang word

i legit thought

meant pdf files

💀

[lucid yoke]

i have a few words in my negative prompt that will always be there to prevent such issues 🙏

[coral ermine]

Any Telegram user can open a bot and type a disgusting request. (there are free tokens for new users)
My code just was sending user ids wrong, that's why I suppose my acc. was blocked.

[lucid yoke]

yes but if your bot didnt have a filter applied blocking those keywords (like civitai and other services has) you could be held liable

[coral ermine]

SD filters such requests itself. We (API users only pass ids to identify who sends requests)

[wicked widget]

yeah i always have (female child, loli:1.2) in my negative

wonder if theres a way you can default negative prompts

to prevent this in the future

[lucid yoke]

there are workarounds but hey just letting you know

[lucid yoke]

dont forget chibi if its a realistic model

[wicked widget]

oh thats a good one, ill be sure to add

i dont use realistic models but good info to know

[lucid yoke]

theres a NSFW filter extension for A1111 but im hoping that it gets ported to swarm

[coral ermine]

Is this offline library you've mentioned? (for filtering)

[lucid yoke]

i dont program but i would imagine youd either have a CSV or other database having banned words

and informing the user those are no no 's before its even sent to stability

[wicked widget]

shit id figure a way out for the users reqyuesting that sort of thing to be logged and sent over to stability and feds

[coral ermine]

Anyways API users just passes ids of these who sends requests. Let SD just handles it itself.

[lucid yoke]

And SD doesnt care about a telegram id doing that and would just ban the api key of it happens too much

[coral ermine]

Moreover there is a special kind of response "content moderation". My fault is I didn't send user ids properly because of a bug.

[wicked widget]

props for you to spot a problematic issue in something you created

[coral ermine]

It happened not much, but because of a bug they couldn't detect who is sending requests and ban requests with specific data in "stability-client-user-id".

[wicked widget]

hopefully SD gets back to you soon

[coral ermine]

Anyways, I hope they will answer soon with email. Let's pray for it.

[wicked widget]

yeah i hope so too

[coral ermine]

Why does HTTP header "stability-client-user-id" exist then?

[lucid yoke]

im not a programmer, but ive seen cases such as these come by before

they end up banning the api key but hey take it up with stablility. was just letting you know to prevent a more serious issue

[coral ermine]

Btw, this incident would not happen if they make this header required.

[lucid yoke]

moderation issues to YOU

so moderation is on you

[wicked widget]

dont add him

random, new user

could be a scammer

happens to often nowadays in here sadly

[lucid yoke]

mostly lack of active moderation (dont blame them)

[wicked widget]

i know :/

[coral ermine]

Zoomers don't like to read. It's too complex.

[wicked widget]

2 yrs ago this was the place to be for SD, it seems the devs, staff and everyone else lost interest

I miss the weekly Q&A we had we SD devs

Ask Emad questions amongst others

[coral ermine]

Anyways, I don't want to moderate anything. I want to pay $ for API and use it. But... it seems I have to.

[lucid yoke]

i mean the alternative is not having it public

dont click on any links from people you dont know 👍

[coral ermine]

Then my service won't grow. Btw different jerks sometimes write disgusting requests to ChatGPT and Dalle in my service, but my API acc. still is not blocked there.

[coral ermine]

I'm so noob in discord. I don't event know how to open PM here. Too complex UI.

I was forced to enter this channel hoping I'll find here support for API unblock.

[wicked widget]

https://tenor.com/view/king-of-the-hill-you-gif-5929698

[normal wedge]

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

call webui.bat

its empty...

[ornate elk]

Then you need to add:
--xformers --medvram-sdxl
To the commandline_args=

That increases the performance and lowers the vram usage.
Then you shouldn't get any issues

[normal wedge]

Mmkay ill try that

[normal wedge]

Hey man just want to let you know the trick did solve the problem. Thanks a lot!! My image generation is a lot faster!!

[ornate elk]

Nice, no problem 🙂

[novel owl]

Hello, I'm coming to you to find out if anyone knows how to optimize Stable diffusion on an AMD graphics card because the performance I have seems lower compared to other cards of the same model that I have.

i have a 7900XT with 20 Go of VRAM

[austere charm]

recently upgraded my gtx 1080 to an rtx 3060 and learned that the generation won't be the exact same as the 1080, which is fine. But I notice that character quality has decreased significantly, such as faces looking weird, characters having wider hips than usual, poses looking weirder, arm placement looking iffy, etc. Are there some general fixes for this?

[ornate elk]

Hey for AMD GPUs you should follow my Guides from here:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides

I'm using a 7900xtx, feel free to ask if any step isnt clear

[novel owl]

okay thx what nvidia graphics card can you compare a 7900 xt and xtx to?

[novel owl]

I chose the ZLUDA solution but now I have a compilation that has been going on for a long time, is that normal?

its good ?

[ornate elk]

Yep thats normal

Compiling will take 15-40 minutes for the first time

[ornate elk]

In terms of ai or gaming?

[novel owl]

AI

[ornate elk]

Hard to compare. Would say like a 3090 maybe

[novel owl]

okay^^

[novel owl]

and its a good perf for my GPU ?

[ornate elk]

Looks good but I can say more when you try a 1024x1024 sdxl image

Should take seconds

[novel owl]

okay i download the model

here is

[ornate elk]

i dont use that tool, can you just generate a 1024x1024, 30 step, euler a image?
prompt: cat,
negative prompt: blurry,
then show me the it/s in cmd or the time in seconds below the image

[novel owl]

wait is not euler a

that it

[ornate elk]

looks good

[novel owl]

what is your it/s with an xtx?

[ornate elk]

nearly the same like 3-3.2max

[novel owl]

okay overclock of the GPU can help to have more IT/S ?

[ornate elk]

yes it could help but i dont know how much, my guess its minimal.
if you overclock, please only use the Adrenalin software for that

[novel owl]

XDDDDDDDD

[ornate elk]

these support scammers...

[novel owl]

that my actual OC

work for all gaming AI work on 3D unity project etc...

[ornate elk]

nice

[vast bay]

@novel owl I advise you create a ticket and contact support team or admin for guidance and assistance. kindly Add me up, I’ll help you to

[novel owl]

look you role stupid man and after you can talk

[ornate elk]

gave him a timeout until admin bans him

[novel owl]

dont wory i juste like trash bad people ^^

[ornate elk]

xD

[novel owl]

the CPU is important for AI or only the GPU when using gpu for AI ?

[ornate elk]

only the gpu

[novel owl]

okay thx^^

[ornate elk]

no problem
for any AMD and Ai related questions feel free to ask me

[novel owl]

also last thing I saw that there were models on CIVIT AI which made GIFs is it possible with automatic 1111?

[ornate elk]

yes with animatediff its possible

but its not that great

[novel owl]

animatediff ?

[ornate elk]

its an extension

for short gifs like 5 seconds

[novel owl]

i have install the extention of animatediff how tu use it now ?

[ornate elk]

you need to download the animatediff models too

[novel owl]

this model is good ?

https://civitai.com/models/1090064/hs-art-anime?modelVersionId=1224226

[ornate elk]

thats a normal model it can work with animatediff but you still need the animatediff sdxl model

https://civitai.com/models/108836?modelVersionId=219642

[novel owl]

ahhh i need a second model to use animateDiff ?

okay

[ornate elk]

here is all info about the extension:
https://github.com/guoyww/AnimateDiff#setup-for-inference

[novel owl]

wat is the argument to make the model folder like this ? --ckpt-dir "D:\IA\model\Stable-diffusion"

[ornate elk]

thats only needed to link the models from a different drive or webui

[novel owl]

yes i want link multiple webui to make a french video to the future for explain how make ai picture and more ^^

[ornate elk]

from my guide:

[novel owl]

its controler or embeding ?

[ornate elk]

?

[novel owl]

for lora is lora for model of picture gen is ckpt and for animatediff model what is ?

[ornate elk]

nothing

its special and needs to be put into the extensions/animatediff/model folder

[novel owl]

ah okay i cant link animatediff model to multiple webui ?

[ornate elk]

nope

only with symlinks

but not needed for one model

[novel owl]

My SSD wana cry

[ornate elk]

you can store everything else on the other drives

[novel owl]

the ssd have 6 month

yep

[formal zodiac]

So how do I make that work? I lost all of my settings when I booted back up

Oh wait

I found it

[midnight kiln]

in reference to the UI state saver? You will have 2 new icons next to the "refresh checkpoint models"

and you can manually save - or set it to autosave a preset every time you gen anywthing

[formal zodiac]

Okay now I am gonna try it again. How do I fix the wacky colors?

So you said I need to find a new model?

[midnight kiln]

that is usually due to using a lora on a model that it's not compatible with

or

if using a GPU that can't do fp16, that can be an issue - if you don't select certain settings

or

[formal zodiac]

I have a 3080

[midnight kiln]

if for some reason the rendering size has been set way too low - models just cant handle it

[midnight kiln]

that has both FP16 and FP8 support - all types. optimal. you don't need any compatibility settings

<---- poor peasant with 1080Ti (but at least I got more VRAM nyah nyah!)

[formal zodiac]

So it might be the model that is the problem

[midnight kiln]

most often it is misusing loras

[formal zodiac]

https://civitai.com/models/404154/wai-ani-nsfw-ponyxl I believe this is the model

[midnight kiln]

using a pony lora on an SD.1.5 checkpoint for example.

[formal zodiac]

Both are pony so idk

[midnight kiln]

Pony base-model - so use any pony loras

you can also to some extent use loras for SDXL (as that's what pony is based on) - it might just not always work perfectly

[formal zodiac]

https://civitai.com/models/549838/phoenix-dota-2-sdxl-pony

This is the lora

[midnight kiln]

you can sort on civitai for loras spesifically for pony (or pony + sdxl)

thats also pony, so that's correct

so thats not the issue then

check these arent out of wack

ps - these are not the settings in hires fix

[formal zodiac]

It is half the resolution that the model's sample photos are

[midnight kiln]

that can work - but usually most compatible to use 1:1 format.

[formal zodiac]

Ah okay

[midnight kiln]

try 512x512 for maximum compatability

[formal zodiac]

Nope that wasn't the issue

[midnight kiln]

and yes, thats normal - you use upscaler to 2x it to 1024x1024 for example. that (or slightly under) will be your max limit you have VRAM for anyway, unless you set the "low-VRAM" or "medium-VRAM" setting (costs speed to free up more VRAM)

check your vae setting. settings -> vae

is it auto, or manally chosen?

[midnight kiln]

forcing a wrong vae on a model can definitely cause these weird color-corruptions

[ornate elk]

have you now updated your webui since the last time?

[formal zodiac]

Oh I forgot LOL

[midnight kiln]

oh, well that's kinda the #1 thing 😛

[formal zodiac]

Where can I find the version info

[midnight kiln]

wahts the easiest way though? I installed just recently and haven't really had the need to check

[ornate elk]

at the bottom

of the webui

[midnight kiln]

anything simpler than giyhub pull?

[formal zodiac]

version: v1.6.0-2-g4afaaf8a  •  python: 3.10.6  •  torch: 2.0.1+cu118  •  xformers: N/A  •  gradio: 3.41.2  •  checkpoint: 4a11de18c7

[ornate elk]

wow thats old

like 1-2 years

[formal zodiac]

I did install this last year and never updated it

[midnight kiln]

shame.. ding shame.... ding

[ornate elk]

whats your gpu?

[formal zodiac]

3080

[midnight kiln]

at the speed things are moving in the image-gen space - that's at least 2 centuries ago

[formal zodiac]

Lmao funny enough I had an issue on my modded 3DS application called Twilamenu and I found out the issue was that I haven't updated the app in 6 years LOL

Wait no at the time it was 5

[ornate elk]

okay then the first thing you do is:
edit the webui-user.bat (right click, edit)
at the second line (which is empty) you add:
git pull
then at the line Commandline_ARGS= you add: --xformers --medvram--sdxl --no-half-vae
then save and close the file.
Then delete the venv folder.
and then relaunch the webui-user.bat

[formal zodiac]

I had an issue during November 2024 and my version was Christmas 2019 LOL

So I definitely have a history of not updating shit

[midnight kiln]

you recommend medram generally CS1o?

haven't tested yet how bad the impact is

and does the 3080 really need no-half?

[ornate elk]

--medvram for gpus with 6gb vram or less.
--medvram-sdxl for 8-11gb vram

[formal zodiac]

Here?

[midnight kiln]

I thiought that was for lacking FP16 or FP8

[ornate elk]

yes but

--no-half-vae is just for the vae process
--no-half is for the lack of fp16 support (only needed for gtx16xx cards)

[formal zodiac]

?

[ornate elk]

yes

[formal zodiac]

Okii

[ornate elk]

git pull

[formal zodiac]

Welp it's gonna take a long time to do that

It's 1.2 GB and going up

LOL

Oh wait

[ornate elk]

it will download the latest torch version (2gb)

[formal zodiac]

That was fast

[ornate elk]

if all is done screenshot the version again

[formal zodiac]

Downloading

[midnight kiln]

For a 1080Ti, can you clarify for me what set of compatibility settings I need? upcast cross attention (GUI), --no-half --full prresicion . I'm not clear if these all effectively do the same or if I should run all 3. I have certain models that freak out and produce gibberish, so I feel like there's something there im still missing

1080 only has fp32 for reference EDIT: apparently I'm wrong, although it sort of runs like a fp32 internally. might still help with VRAM though

I believe..

[ornate elk]

A 1080ti only needs --xformers --medvram-sdxl --no-half-vae

Then you have the best possible performance

[midnight kiln]

super!

[ornate elk]

Also forge webui could be a bit faster too

[midnight kiln]

not even the upcasting setting?

[ornate elk]

Nope

[midnight kiln]

then I've really had everything wrong lol

no wonder stuff was messing up

[ornate elk]

I used a 1080 a long time before my new GPU

[midnight kiln]

Ah, so you know my pain 😛

Really considering getting a 3090 if I can find a scoop on it. That would be nice.... 24GB bites lip

[ornate elk]

But with the settings I mentioned you can easily use sdxl with an okayish speed

[midnight kiln]

speaking of, what is the --sdxl switch you showed above? never seen that

[ornate elk]

3090 is quite rare and used ones mostly are broken caused by mining

[midnight kiln]

yea, wont buy one unless i can know it's past somewhat reliably

[ornate elk]

--medvram-sdxl applies --medvram but only for sdxl models. So that you dont lose speed for 1.5 models and get the better speed for sdxl

[midnight kiln]

cool

[ornate elk]

If you really used --no-half --Precision full you slowed your self down by alot

[midnight kiln]

I think I've actually been using mostly just the upcast in the GUI. It's just been bothering me that I had weird issues

[ornate elk]

Ah okay

Let me know if you get any issues again

[waxen pasture]

hey, in the webui thingie for some reason i csnt see all the loras i placed in the loras folder

i dont get this thing of the devil

[midnight kiln]

did you hit the refresh button? or restart the GUI ?

they should be showing up then

I guess you can check that the extension is correct on the files - just so you havn't confused something else for a lora

[ornate elk]

They will only show up when a compatible Model is selected

[midnight kiln]

Oh really?

didn't know that

[waxen pasture]

yeah!

[midnight kiln]

read what CS1o said - he knows his stuff

[waxen pasture]

Whaa, so some models are incompatible?
like stable diffusion models or checkpoints

[midnight kiln]

those are kinda the same thing no? used the same at least

but yes, you dont want to use a pony lora on a checkpoint running SD1.5 for example EDIT: sorry I meant to say the opposite, but neither would work since Pony is based on SDXL, not SD1.5)

But sometimes you can go the opposite way, where using Pony (based on SDXL), you might still get SDXL lora to work

always get the spesific version if it exists. Note that on civitai, a download page often has many options (near the top). Its easy to overlook. for lora this can often show you the same lora trained to fit different checkpoints

[midnight kiln]

You don't happen to know what the correct setup of torch, torchvision, torchaudio and xformers is? I am having such a nightmare sorting out the dependencies after enabling --xformers

[waxen pasture]

oh, they are?

[midnight kiln]

well... like, pont is just SDXL, but further customized - and it became sort of it's own thing

[waxen pasture]

ok so different loras will appear based on the checkpoint i use?

[midnight kiln]

but SD1, SD1.5 and SDXL are like the official starting points. Everything else grew from those parents

[waxen pasture]

and checkpoints are the same as models?

[midnight kiln]

a model is a lot of things - including lora. its a very broad term

i think checkpoints is accurate to describe both the official ones and the customized stuff. People may say "the base model", but like I say, model is super broad

all thse things are models:

don't quote me on this, but I think the term is generally used to refer to neural networks - but as you can see they have very widely different uses. a pose is what it sounds like. a detection is auto-finding a thing in an image, like an eye

and no shortage of base models either

Although not all are so heavily used. just be aware of them 🙂

[fair oxide]

i have amd 7900xtx can it run ai img gen orrr no because of the cuda thingy?

[safe inlet]

you can, there is info in pinned messages here

[midnight kiln]

Does SD/automatic1111 make use of cuDNN?

Does someone know the solution for this?

------transcribed windows runtime error box--------
The procedure entry point
??0Error@c10@@QEAAV?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@@Z could not be located in the dynamic link library
C:\stable-diffusion\stable-diffusion-webui\venv\Lib\site-packages\xformers_C.pyd

EDIT: Nevermind, fixed finally. If anyone has the same issue just Pm me.

[mystic basin]

My automatic1111 did a git pull automatically and now I get errors, I don't know why. ( I need help)

[fair oxide]

why not use that webui thingy

I have a rtx 4070 super and image generation is still pretty slow at 1024 by 1024

are there any ways to speed this up without reducing quality a noticeable amount

[mystic basin]

i use automatic1111 web ui

[mystic basin]

redownloaded everything and i get this error now

• I also get the error I was getting before redownloading everything

[ornate elk]

Go for my Zluda Guides. First link of the pinned messages.
I have the same GPU and no issues

[mystic basin]

shouldn't have redownloaded everything

[ornate elk]

What's your GPU?

[mystic basin]

RX6850mxt

I thought redownloading everything would fix the OSError I had but it did not and now I have another error + the one I had before trying to redownload automatic1111 with ZLUDA

[ornate elk]

Did you followed my guide from scratch?
Maybe you downloaded the wrong zluda version.
Rocm hip SDK 6.1 needs the 3.8.5 zluda.

Did you reinstalled hip SDK too? That would mean you also need to redownload your GPUs library files

[ornate elk]

If you use Auto1111, edit the webui-user.bat
And add --xformers to the commandline_args

[mystic basin]

I did not, I thought since I already downloaded it before, I only had to do the git clone part and modify the web ui user file. (I already have the path and I already modified folder in : C:\Program Files\AMD\ROCm\6.1\bin\rocblas

[ornate elk]

Try delete the .zluda folder and relaunch
If that doesn't help then copy over all the files from C:/zluda into stable-diffusion-webui-amdgpu/.zluda

[mystic basin]

alright

it didn't work

tried both

[ornate elk]

What's your python version?

Try updating it to 3.10.11 64bit and then delete the venv folder and relaunch

[mystic basin]

alr

it doesn't find python anymore

[wicked widget]

ur not installing python through windows store are you? thats a big nono

[ornate elk]

Open up a cmd and type
Python --version

[mystic basin]

doesn't find it

i have the path as a variable

[ornate elk]

Then install 3.10.11 64bit

And delete the venv folder

[mystic basin]

that's what I did

[ornate elk]

Then python --version should work

If not it could be a third party antivirus broke it

[mystic basin]

when I do where python, it find where python is

[ornate elk]

And and where is it?

Can you screenshot it

[mystic basin]

this isn't the path that is in the variable thing, so I don't know what this path is

[ornate elk]

The top one is wrong

Thats from ms store

[mystic basin]

I saw that

I don't know why it says that, maybe installed it once

[ornate elk]

Uninstall every version

Via System control AND trough windows apps

[mystic basin]

nope, it send me to the microsoft store

i deleted every python folder

other than the one needed

i kept python 3.10.11

still doesn't find it

when I use the windows app removal thing, it doesn't detect any of the python version that i deleted the folder and can't uninstall

so they should be gone

[waxen pasture]

i see
what a pain!
why must people be so confusing!

[wicked widget]

Whats confusing you

?

[mystic basin]

managed to make the web-ui user find the python path by modifying the folder and adding the path (won't fix the problem of the cmd not finding it but it's not something i need to worry about right now)

[austere charm]

is there a fix to character generations not looking right anymore? Upgraded my GPU and now torso's are too long, bodies are too wide, etc. I've been utilizing the same settings as my previous gpu

[ornate elk]

You can add the Python path to the webui-user.bat

[wicked widget]

show me ur settigngs

and tell me what gpu/vram

[mystic basin]

that's what i did

i'm waiting for the install to finish

[austere charm]

sure, what's the best approach for sharing the settings? Also RTX 3060 12GB VRAM

[wicked widget]

taking a screenshot of the webui

with ur last generation

and perfect thats my card

[austere charm]

this one right?

[wicked widget]

not quite, when youre in the webui generating content

show me those settings

so i can see what model, prompt, etc

since u have 3060 tho u can add --cuda-malloc --cuda-stream to that tho after xformers

itll help

[austere charm]

this okay?

[wicked widget]

perfect gimme a sec to look

[austere charm]

for sure, thank you 🙏

[mystic basin]

it find python correctly and launch now but I still have the error

[wicked widget]

everything looks fine, what seems to be the issue exactly? can you show me examples

[austere charm]

I find that bodies typically aren't as good as they were on my previous gpu (1080), issues ranging from poses just looking iffy, hands looking worse, bodies being either too wide or too long, on top of essentially being unable to recreate images i've generated on my old gpu

[wicked widget]

one thing i do notice is you are applying too much weight (word:1.2) is considered really high but not breaking anything, cfg scale is abit too high, i personally dont use clip skip as i dont find it helpful but personal preference, everything else seems okay, pony needs all those extra token bloat to be good which is why i tend to use Illustrious models

thats a 1024 for pony

should be fine

[austere charm]

that does look pretty good! I'll mess around with those settings you mentioned, thank you so much Ellie! 🙏

[wicked widget]

np gl

ping one of us if need help

me voluntelling cs
ok ping ME

😛

[ornate elk]

Hes not using forge

[wicked widget]

o

should probably use forge

tbh

[ornate elk]

Oh wait

He is

My bad

[wicked widget]

XD

tired?

[ornate elk]

Yep

And at work xD

[wicked widget]

[ornate elk]

Have to fix User network drive connection

[ornate elk]

No wait hes not.
That was your Screenshot xD

[wicked widget]

XD

[wicked widget]

you should porbably also look into downloading a1111 forge

itll run better on our card

has some good built in extensions

[austere charm]

for sure, i'll give it a look, thanks! 😄

[wicked widget]

np

[tidal thunder]

Hello guys, can i get help here about the forge version?

[wicked widget]

whats happening

[tidal thunder]

I'm getting this error: https://pastebin.com/xDL9xjTY

while trying to run controlnet (ip-adapter)

on xl model in stable diffusion forge

[wicked widget]

something to do with the ipadapter yeah, show me those settings pls

[tidal thunder]

not only ip-adapter work's not normally, reference is half working...

[wicked widget]

hit pixel perfect, and then enable preview, does the preview work? if not, its the preprocessor messing up

[wicked widget]

hit the lil explode button next to preprocessor

(i honestly never used ipadapter and using general controlnet knowledge to attempt to help)

[tidal thunder]

i've tried to reinstall venv folder, do something to code with chatgpt, do something in start arguments, but it doesn't help

[wicked widget]

is there another preprocessor thats ipadapter? id need to mess with it more myself so i can help better in the future

[tidal thunder]

there's 2 more, yes, i'll try it now

[wicked widget]

try those others ones ya

typically when something breaks in controlnet is a mismatch of preprocessor and controlnet model

[tidal thunder]

still nothing.. RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument tensors in method wrapper_CUDA_cat)

[tidal thunder]

but every model is.. broken?... how that could be?

[wicked widget]

which base model are you using, an sdxl based?>

[tidal thunder]

DreamshaperXL turbo

but even default dreamshaper isn't working

[wicked widget]

ok try this,
keeping the ipapdater preprocessor, download https://huggingface.co/xinsir/controlnet-union-sdxl-1.0 this

use this controlnet model in place

with the ipadapter preprocessor

this union controlnet is like a super controlnet, knows all the controlnets into one

[tidal thunder]

i'll try now

[wicked widget]

canny, lineart, etc

[tidal thunder]

wow, a new error

https://pastebin.com/yug0cit9

[wicked widget]

what resolution are you trying

[tidal thunder]

648x648

[wicked widget]

way too low for xl

1024 is the standard

that couldve been the issue all along

[tidal thunder]

1500x1500 okay? it didn't work AttributeError: 'dict' object has no attribute 'shape'
'dict' object has no attribute 'shape'

[wicked widget]

try 1024x1024

[tidal thunder]

no

didn't work

[wicked widget]

im unsure what could be the issue unfortunately, yhopefulyl someone with more controlnet knowledge can help

[tidal thunder]

Hmm, okay, anyway, big thanks for the help! i've couldn't figgure out without your help about ultra model for controlnet

[wicked widget]

wish i couldve fixed it but controlnet in general is fickle and thus why i dont mess with it too much,

the union model is really nice however, try it with another preprocessor, wondering if just ipadapter being dumb

[tidal thunder]

non of the preprocessors are work

[wicked widget]

hmmmmmm. im starting to think the installation of the ui is borked.
can you completely uninstall the ui and try a more optimized one like a1111 Forge? it has controlnet built in aswell

[tidal thunder]

i've tried to reinstall

[wicked widget]

😦

[tidal thunder]

i'll try to do something with union models

[ornate elk]

Forge is a bit tricky with controlnet

[fair oxide]

I already did this
Is there anyway to speed it up further?

[manic flume]

How do you get illustrious checkpoints to work with a1111? Ive got the base model Illustrious XL (also SDXL and Pony working) and with illustrious I just consistently get like all white images

The first bit of generation it seems to try and make something but then it all just becomes the same white color

Issue seemingly solved

Apparently removing all my negative prompts got it working. Not sure if there was a specific prompt causing the issue but its weirdly just for Illustrious models and not SD1.5, SDXL, or PDXL

[ornate elk]

Use DPM++ 2m or Euler a

As sampler

And when using hires fix set hires steps to 10

[manic flume]

Ah, it was just caused by going over 75 tokens, i dont get the same issues there with other models but yeah... Found an extension that fixes the issue

[fair oxide]

pro max is probably a better choice https://huggingface.co/brad-twinkl/controlnet-union-sdxl-1.0-promax/tree/main

[wicked widget]

pro max uses more vram

if you can run it then ya

but i recommend the base first when idk the specs

[fair oxide]

not quite, they both use pretty similar memory footprint, except promax has more features than the other

oh drats, i'm posting images in tech support chan....

[tidal thunder]

Find a sollution!!!

i just change UI

from XL to sd

LMAO

[wicked widget]

oh wow lol

nice catxh

[lapis wind]

hi I'm getting this error after I do a second generation

[ornate elk]

What's your GPU?

And which settings did you used?

[lapis wind]

rx 6600

[ornate elk]

If you used hires fix you need to enable the never oom option at the bottom of txt2img

[lapis wind]

i don't use hires fix

[ornate elk]

Ans which resolution, steps, and model did you used?

Also can you show the full cmd log?

[lapis wind]

animeModel_amorev2PDXL, 512x768 and 30 steps but before I formatted my pc everything was working fine, now it's giving me these errors again

[ornate elk]

Okay then a Full cmd log pls

Also if you have 16gb RAM you need to increase the Windows Pagefile

[lapis wind]

I set the pagefile to 16GB

[tidal thunder]

help me please, i need to somehow turn on (permanently would be great) gpu weights (mb) in sd profile (no, i can't switch to xl UI)

[lapis wind]

I'm regenerating the image and I'll give you the logs

[rugged summit]

does anyone know why on forge it keeps giving error cuda out of memory? it didn't do it before and low/med vram isn't working anymore. It keeps trying to allocate 12gb of memory on my 8gb card

[fair oxide]

Alr thx will try

[ornate elk]

You need to set it to 16000min and 24000max

For C

And disabled for any other drive

[ornate elk]

You need to click XL on the left

[ornate elk]

Which model and settings do you use?

[ornate elk]

You need to upgrade to python 3.10.11 also you should switch to a newer hip SDK and zluda version

[lapis wind]

I should just delete the folder venv from Stable Diffusion ?

[ornate elk]

I thought you installed everything fresh?

Deleting the venv won't help

Only if its old

[lapis wind]

I'm deleting the folder venv for the new Python version

[ornate elk]

Yep that's needed

[rugged summit]

using sdxl 896x1152 generation is fine but i cant upscale

[ornate elk]

whats your gpu?

enable the never oom for tiled vae at the bottom of txt2img and then set hires steps to 10 and it should work

[rugged summit]

rx 7600 using zluda

[rugged summit]

i’ll try

[rugged summit]

didnt work, pc froze for a bit and output changed from trying to allocate 12gb to 4gb

also tried img2img, other upscalers in extras work but latent doesn’t sadly

[ornate elk]

Hmm then you should lower the resolution

Or set upscale by to 1.5

[formal zodiac]

So uh, I tried updating the webui thing and now I get this error

@midnight kiln

[ornate elk]

Your missing a -

It needs to be --xformers

[formal zodiac]

Oh

Now there's this

I think I need to put a space?

Okay now it's this

[ornate elk]

Can you show the full cmd log?

[tidal thunder]

i can't because of the controlnet working only when ui enablen as sd, so i need to change gpu weights in sd ui

[wicked widget]

--medvram-sdxl

got an extra -

[ornate elk]

Then you need to download the controlnet models for sdxl

[tidal thunder]

that's long problem and i found only 1 solution it's to change ui from xl to sd, and i need to change gpu weights on it

[ornate elk]

Like Ellie said, its --medvram-sdxl not --medvram--sdxl

[formal zodiac]

Oh

[ornate elk]

Can you screenshot your controlnet models?

[tidal thunder]

yes, sec

[formal zodiac]

Okay now I think it's working

Okay it opened but there was an error

[ERROR]: Config states D:\animefull-final-pruned\stable-diffusion-webui\config_states\civitai_subfolders.json, "created_at" does not exist

[tidal thunder]

https://pastebin.com/LmZkr0Sa

[ornate elk]

Go on the extensions tab and click check for updates.
Let it load and then restart the webui

[formal zodiac]

Do I need to make a folder?

Oh okay

[tidal thunder]

and the strangiest's thing, is that ip-adapter, revision and sometimes reference don't work at all

[ornate elk]

What does the cmd shows?

[ornate elk]

What's your GPU?

[tidal thunder]

rtx 2070 super 8 gb

[ornate elk]

Okay maybe its to much for the GPU. Have you tried an other preprocessor or model without clip?

[tidal thunder]

at sd ui all work's fine, and on xl ui, it's... don't work

[tidal thunder]

didn't work too

it's depth

[ornate elk]

How big are the controlnet models?

[tidal thunder]

600~ mb

[ornate elk]

Hmm that's okay

[formal zodiac]

Okay so updating the webui didn't fix the generation @midnight kiln

[ornate elk]

What error do you get?

[ornate elk]

Can you try a lower resolution? Like 512x768

And are all your controlnet models for sdxl?

[tidal thunder]

on depth or ip-adapter?

[tidal thunder]

yes

[formal zodiac]

Okay so it DOES generate but it gets more and more deepfried with each update

[formal zodiac]

Like here

[ornate elk]

Depth or openpose

[formal zodiac]

I have a theory that the Lora I used is incompatible with my model

https://civitai.com/models/404154/wai-ani-nsfw-ponyxl

[ornate elk]

And when you try without it?

[formal zodiac]

This is the model

Hang on lemme try

[tidal thunder]

depth now start's normally (i turn off kohyahighres)

[formal zodiac]

Yeah it was the lora

[lapis wind]

I update python and hip SDK, now it shows me these errors

[midnight kiln]

Hey all - anyone know if this is an issue that needs attention? (from terminal log):
"AssertionError: C:\stable-diffusion\stable-diffusion-webui\models\Stable-diffusion\ponyRealism_V21MainVAE.safetensors is not a safetensors file"
The files comes straight from civitai, and it's a popular checkpoint. I checked the hash to ensure it's not corrupted.
Would this sort of thing be a potential security risk?

[tidal thunder]

why i'm getting red/purple image when trying to do somethign with reference (exception is reference_only)

[formal zodiac]

Here's the result without the lora, but I wanted the lora to be more accurate to what I was looking for

So Stigma said that I need to find a model that is compatible with the lora

https://civitai.com/models/549838/phoenix-dota-2-sdxl-pony

I don't get it because both the model and lora are pony

[midnight kiln]

Yes. The lora is pony, so are you using a pony-based checkpoint?

[formal zodiac]

Right here

[midnight kiln]

then it should be good

same issue as yesterday? let me read your posts

[formal zodiac]

Yes same issue as yesterday

[midnight kiln]

what checkpoint did you use again? I can just test it on my side real quick to see if I can replicate it

that will narrow doen the issue

[formal zodiac]

Again, it's the one that I linked

[ornate elk]

Have you adjusted the Pagefile like I said?

16000min and 24000max and then reboot the PC

[lapis wind]

I set to 16000 min and max

[ornate elk]

Thats not good

[midnight kiln]

ok im grabbing it now. But I do make a note that there is a page from the same author, same model name, that goes all teh way up to version v12. you seem to have v6. You may consider updating. Not that newer versions are always better in all ways, but they tend to be overall improvements:
https://civitai.com/models/553648/wai-ani-hentai-ponyxl
https://civitai.com/models/404154?modelVersionId=1295881
(note to other readers - NSFW models, click at your own discretion)

I will test the same setup you use to recreate

[wicked widget]

thats awful, raise it

[formal zodiac]

Downloading

[lapis wind]

I've now set it to 16000 min and 24000 max

[formal zodiac]

Saw this on the page, what does this mean?

[midnight kiln]

Adetailer is an extremely useful extension

it basically runs through an image - detects faces, eys, hands ect. and gives them extra passes - optionally with higher resolution

[formal zodiac]

Ahhhh

[midnight kiln]

as these areas are those that needs most help - you can focus on them, leading to far less broken faces that ruin an otherwise good image

[formal zodiac]

Ahhhh okay

[midnight kiln]

you can run it on every image gen, or use it to touch up after in img2img

It's a must-have honestly. Let me know if you need help setting it up

[formal zodiac]

Okii

[wicked widget]

raise the min

[formal zodiac]

Lemme get the updated model in there

[midnight kiln]

You can also do other nice things with it - like giving eyes extra passes with a spesific eye-lora. Or using the optiona prompt in adetailer to say the eyes should be blue for examples. It is also extreeemly useful to give emotions to characters. I usually generate them fairly neutral - and then leave adetailer's face-tab to mold the expression into what I want - make them look a certain direction ect. ect.

because then you can easily play around with the faces after a long gen. Unless you render them in super-resolution it will be very fast as it only needs to run on a relatively small area

molding just the right expressions into faces is the stuff that takes things to the next level

you can even add more detectors for other things to search the image for. There are plenty

[formal zodiac]

Oh so it WAS the model not being updated that was the issue!!

[wicked widget]

typically is

[lapis wind]

thanks it's worked

[rugged summit]

this worked, is it normal for latent to be so slow? around 30s/it while generating is 3s/it

[ornate elk]

Try Resrgan4x if its faster

But even 3s seems a bit slow

[rugged summit]

do you think so?

[ornate elk]

Depends on the model and resolution but without upscaling it should be faster I guess

[ornate elk]

Which hip SDK and zluda version do you use?

[rugged summit]

6.2.4 and latest zluda

[ornate elk]

Okay then try add --cuda-stream to the webui-user.bat
Also how much ram do you have?

[rugged summit]

16gb and using an 8gb card

[rugged summit]

im gonna try adding that to my next generation

[ornate elk]

Then you need to increase the Windows Pagefile

That can speed up the sdxl generation

[rugged summit]

to what should i change it? its currently at 8700mb

[ornate elk]

To 16000min and 24000max for C

And disabled for any other drive

[rugged summit]

i did everything and now im currently on a best of 1.54s/it

its upscaling with esrgan 4x at around 20s/it

[ornate elk]

Can you try a 1024x1024 sdxl image with 30 steps. No hires fix.
Sampler Euler a

[rugged summit]

around 3.80s/it

[ornate elk]

how long in seconds?

[ornate elk]

have you restarted the PC after the pagefile change?

[haughty halo]

hello, i use ComfyUI on Arch linux (hyprland). Can i find help here with some technical questions?

[coarse hinge]

hi, SPAR3D has been generating flat 3D objects that lack the correct topology. Any tips on how to improve this?

[austere charm]

if anyone else had a similar problem to me, i just deleted my checkpoint and vae and re-downloaded them and now everything looks great, I honestly should have attempted that way earlier 😆

[limpid osprey]

Like?

[fair oxide]

apple

||if u sign up with gieco today, u could be saving 15 percent or more on ur car insurance||

wut was that

xd

[ornate elk]

Support scammer

[fair oxide]

oop

im glad i joind this server and was pretty open about my prompts and sent screenshots of my terminal here

made it easy to get back wut i had for reinstall xd

[ornate elk]

Your camu right?

[fair oxide]

ya

cant believe u remembered me ;o