🤝｜tech-support

[shut apex]

Ah wait...

It's RAV !

I think it's RAV that's been pissing me off from the start and deleting my file

[vocal burrow]

rav ?

[shut apex]

RAV Endpoint Protection

He did this when I launched the bat file, after deleting venv

[vocal burrow]

never heard of this RAV before tbh.... Seems like it s "one of those" softwares that install by default when you install other stuff and don t uncheck certain checkboxes during installation

those kind of software automatically go on my naughty list. I would uninstall it if I were you tbh

you don t need avast + rav + probably windows defender

but that s another topic.

[half island]

Uninstall the avast and rav (whatever that is) bloat and just stick to windows defender. It's free too lol.

As it appears rav has a lot of false positives and isn't accurate, so if anything, rav appears to actually be the danger, not what it think is danger

[shut apex]

After desinstalling RAV

webui.bat

[vocal burrow]

it's still working

seems fine so far

[shut apex]

So am I finished? Or should I do something else? (I lost track with all this 😅 )

[vocal burrow]

wait for it to finish installing

[brittle lantern]

hey i need some help running a RTX 6000 and has given me this it usally stops flux gens at 50 or 70% on both forge and comfy

and would output but it isnt finished steps

also the step count in terminal doesnt show

as u can see

doesn't go from 1-30

[shut apex]

I have this looping now

[vocal burrow]

relaunch it, it should be fine afterwards

can happen with the first run

also remember that with zluda the first gen is gonna take a looooooong time because it has to compile zluda stuff in the bacground.

[shut apex]

I am here now

[ornate elk]

Try use a smaller flux model. Q8 GGUF for example

[ornate elk]

Looks good

[shut apex]

What should I do now?

[brittle lantern]

ah ok

so 24gb vram isnt enough or sum else?

[ornate elk]

Check the first link of the pinned messages. There the guides a updated, but the other ones work too

[ornate elk]

If you try to use the 23 GB flux model with a 7900xtx it wont work

Q8 gguf works good

[shut apex]

Where do I get this?

[gusty solstice]

Hi!
Is burned cable still an issue with the RTX 5090?

[vocal burrow]

it's on the last screenshot you've sent

[ornate elk]

Provide as much info, and I can check when I have time.
GPU, driver version, webui, complete cmd log.

[shut apex]

I mean where do I put it?

[vocal burrow]

in your internet browser

[ornate elk]

Yep as its a hardware failure build buy nvidia. There is no fix until they redesign the 12v power connector.
But seasonic showed a new power supply that can detect heat and irregularities and switch off

[shut apex]

I think it's this right ?

[ornate elk]

Kohya won't work with AMD. On my Github page you find a guide to setup OneTrainer with Zluda instead

[ornate elk]

It seems like you still need a base model

You can get them on Civitai.com

Provide me as much info and I can check if I have time.
GPU, driver version, webui, complete cmd log

[shut apex]

Ok. Thanks for your help. I've been trying to install this for 4 days XD. Are there any details I need to know about how it works?

[ornate elk]

Hover over the buttons/text for short infos

Positive prompt box is what you want to see in the Generated image.
Negative prompt box is what you dont like to see in the image

[brittle lantern]

Should Nvidia be more optimised or am I tripin I though RTX 6000 24gb would be fine for full models

[ornate elk]

Thanks, its very nice so far!
Let me know if you still need help

[ornate elk]

Oh sry I was confused xD you said rtx 6000 and I thought of amd RX 6000

A rtx 6000 should have no problems

But the vram still should be not maxed out

[brittle lantern]

Yeah I figured but I’m getting that

[vocal burrow]

in two years or so whenever they get released (rtx6000 :p)

[brittle lantern]

And it stops after like the bar goes to 80%

Or on forge 50-65% it says moving models to GPU and stops gen

[north zodiac]

GPU: AMD Radeon RX 6750 XT (12GB VRAM)
Driver Version: AMD Software Adrenalin 25.4.1
WebUI: https://github.com/lshqqytiger/stable-diffusion-webui-amdgpu-forge
It's also worth mentioning that I just installed the newest full build of Python instead of the earlier version stated in the video tutorial I followed. It seems to have produced a new issue.
Will add the CMD prompt later. Cheers again mate.

[brittle lantern]

Chroma works fine

[brittle lantern]

But other Flux models doesn’t

I’m Chroma is a aggressive fine tune

[ornate elk]

Ia the chroma model the same size as the flux one?

[brittle lantern]

Chroma is flux schnell but like heavy fine tune it has its own comfy nodes

See

And it stops I’m using GGUF Q8

Oh wait

Might be the sampler

Seems to work now fine

It works on Euler normal

Hmm

Tf

Guess I figured it out

Odd

Very

Odd

[ornate elk]

Okay very strange

[brittle lantern]

DPM Adaptive and DDIM and schedular

Was a issue

Even tho the tuner for RayFlux

Said that’s recommended for the model

[ornate elk]

Ah yep ddim doesnt work

[brittle lantern]

Ah for both comfy and forge it doesn’t work Ty

Now I know it ain’t the GPU being “old”

[shut apex]

how do I solve the problem "ValueError: Failed to recognize model type!" ?

[ornate elk]

What's the model you try to use?

[shut apex]

Ah wait

I took a Lora instead of a checkpoint. maybe that's it

[ornate elk]

Yep that won't work

[fossil dew]

Haven't ran into issues yet, aside from it lagging out alot, but I haven't gotten that error again, I'll lyk if I run into any issues~ ❤️

I would however like to be reminded of what it is that's used with Stable Diffusio nt o make moving images 🤔

[ornate elk]

Alright! For auto1111 that was AnimateDiff

[fossil dew]

Ah, Im using forge now 🙂

idk if that's like- different per say, i don't follow alot of this

[ornate elk]

I think AnimateDiff doesnt support Forge

[fossil dew]

huh I swore it had a different name

used to be a big part of the server 🤔

[ornate elk]

Stable Video?

[fossil dew]

hmmm doesnt sound right either, im going to go try and dig and see if I can find it

Deforum

[ornate elk]

Ahh true

[fossil dew]

huh so theres none that work with Forge?

• Stable Video Diffusion
• AnimateDiff for Forge
• Deforum for Forge

so I guess maybe there are options

[ornate elk]

I dont think they will work with forge

[fossil dew]

womp womp

Okidoki n_n

[sudden pike]

sad. Can i train Illustrous models with it?

[runic fjord]

ive been using a1111 for over a year and in the last month or so my generation times have gone way up. i noticed that it like to get hung up right before swtiching to hi res fix and right before the end of the batch generation

it used to be like ~1min for a 768x960 with hi res set to 2x and now its like ~1.5min to ~4min

itll just get stuck right before the end or right before going to hi res fix

[sudden pike]

ya know how i can fix "out of memory" on Onetrainer?

16gigs of VRam isnt enoughj lol

[normal vector]

Hi i've been wanting to use Stable diffusion for quite some time and i've been getting alot of trouble installing, it's either stuck with python or with the torch thing, i'm relatively new to these things would love to have some help

[sudden pike]

@pure tusk

u will be banned soon

im 99% sure

doesn't change a thing

[shut apex]

An other one ? 👀

I don't know why but he doesn't recognize my Loras. It must come from there

[sudden pike]

my coding knowledge is dumber than toast

[vocal burrow]

unlikely, most likely you're just mismatching loras and your checkpoint

you can't use SD1.5 based loras with sdxl models

or sdxl with sd3, etc

[shut apex]

I was talking about the ban. The screen had nothing to do with it XD

[shut apex]

Ok I will check that

[ornate elk]

Lower the batch size to 2 or 3. default for sdxl loras in OneTrainer is 4 and uses 18gb

[low needle]

Hi everyone, I wanted to suggest a quality-of-life feature for txt2img (or img2img) regarding prompt control in multi-character scenes.

Many users create scenes with two characters (e.g., male and female), and it becomes very difficult to control poses, skin color, and attributes separately using a single global prompt.

Currently, we can use the Regional Prompter extension, which allows positional prompt targeting (e.g., [0][0] for left character, [0][1] for right character), but it’s not native and requires a learning curve.

👉 My suggestion: Add a built-in dual-prompt (or regional-prompt) feature in A1111, so users can define separate prompts per character from the start — directly from the UI, similar to how prompt-matrix works.

This would drastically reduce prompt conflict (e.g., hand position, skin tone) and help users who work with multi-subject renders.

Just wanted to hear thoughts from the devs and others. Has this been considered for native integration?

Thanks in advance 🙏

[sudden pike]

k ill try it 😄

[sudden pike]

doesnt work

it somewhat takes 16+gb peak and yeah

it uses Zluda, can it be that my Driver is to new?

[ornate elk]

25.4.1 or 25.3.1 will work. 25.5.1 has some issues

[sudden pike]

idk. even with a batch of 2 it wants more vram than i have

and i have 16 gigs

[ornate elk]

Hmm strange

[sudden pike]

it uses less in general

but then suddenly it uses alot and crashes

this was with an Batch size of 2

[sudden pike]

update. older version works

[ornate elk]

Older driver version?

94gb ram damn xD

[sudden pike]

25.4.1

25.5.1 makes problems. probably cuz Zluda

[sudden pike]

yeah i kinda need it for Music production

[sudden pike]

u know how i can increase the steps?

it just does 8 steps per Epoch

[ornate elk]

In the ui of OneTrainer, but rn im on vacation and can't look at it

[sudden pike]

also itsn ot normal that it takes like 2 Minutes for 20 Epoch right?

and that i dont see any "Training dataset tags"

[ornate elk]

Idk, I only trained one lora so far

I linked a easy lora guide at the bottom of the OneTrainer install guide

[sudden pike]

AHHH found it

[lime pivot]

Do you guys know how I can fix fucked up hair and nose and mouth disappearance when using tiled diffusion to upscale ? Currently im using photo editing to fix the issue but when i have 20+ images in my dataset and each one I have to photo edit by hand, its tiring

[iron lantern]

Can you make gifs ... in stable diffusion with images ... from pony models ?

[lucid yoke]

Using video models yes but you need high end hardware for that (+12gb vram Nvidia)

[iron lantern]

would a 2080ti be ok ?

[lucid yoke]

Maybe the wan 1.3b model?

Unsure about the 14b

You'd need comfy/swarm for it though

[iron lantern]

forge ui no good ?

[lucid yoke]

It can't run video models

[west mural]

Hi
I am wanting to look into how to get SDXL running on an 6900xt, ideally using A1111 (can also use a different one, just prefer not amuse since it is supposedly censored/filtered ?)

Thank you in advance

[lucid yoke]

the UI isnt censored, just the Models can be

in the pinned messages theres a AMD guide

granted Cs1o who wrote them is on vacation so your on your own for the most part

[west mural]

I gotcha, it seemed to be kinda weird to use when I looked at it yesterday (amuse), but I will look into the guide, thank you!

Which webui is a recommendation or is just automatic1111 pretty solid ?

Also gonna be a dumb question but how do I ensure I got SDXL ?
Or is that dependent on the models I grab ?

[lucid yoke]

If your gonna use SDXL i recommend using Forge

on https://civitai.com/ and then then under the models tab you can filter for the "Model" and then for "checkpoints"

if you want anime style images i recommend illustrious

[west mural]

Does it work fine with automatic or should start from the top with forge ?
Since I intent on using mainly SDXL

[lucid yoke]

It works with a1111 but it's slower and a1111 is not maintened

[west mural]

It errored out and gave me memory allocation issues, installed forge and it just worked instantly

[lucid yoke]

Forge too lately but it's the best amd option

[west mural]

Thank you very much for your help tho!

[lucid yoke]

You used the zluda installation?

[west mural]

Yes

[lucid yoke]

Yee, your gonna have some weirdness with upscaling iirc but tiled upscaling will prevent out of memory

[west mural]

Doing 896x1152 native with out 2s/it, also not constantly crashing out of memory which I did a year or so ago with directml

Still seems to constantly run out of vram, so gotta look into that
i like how I say it runs fine and then just dies

[visual bane]

I’m researching methods to build a 2D multi-garment virtual try-on — for example, a person is wearing a shirt. The user selects a suit , and the system should generate a try-on result the suit over the same shirt that respects the order, interaction, and layering of these clothes so is there approach anyone can suggest me?

[fossil dew]

@ornate elk - You said lmk if I have any more issues 🥲

it's done this 3 different times now. It forks for a little bit, but then does this.

I have little memmory so you suggested Forge 🙂 (which has been working great)

[ornate elk]

With which settings do you get this?

[fossil dew]

uh

basic settings I suppose

I dont mess with it too much

[ornate elk]

Have you Never oom for tiled vae enabled?

At the bottom

[fossil dew]

no I can enable it tho

[ornate elk]

How much vram has your gpu?

The never oom setting can help prevent out of vram issues.
But if your error is again related to normal RAM then there is not much you can do other than restart the PC, close any other Programm or getting more RAM.

[fossil dew]

Vram is 8

👍 yeah i havent reset in a muinet ill do that soon

[ornate elk]

Ah okay then try enable the "never oom for tiled vae only" when upscaling anime / semi realism images

[whole sparrow]

Currently trying to delete my stability matrix folder from my main os folder. I copied it into my external hard drive but it keeps saying i need admin permissions to delete it

[pale garden]

SD.NEXT is the recommendation for AMD Windows

The UI is horrendously bad though

[ornate elk]

I would recommend anything other than SDnext

Its not user friendly

Auto1111 or Forge with Zluda is the Go To

[silk bronze]

Is there a problem with ComfyUI and AMD? I was running it before I switched to Nvidia, didn't seem any slower than the other options.

[silk bronze]

Regardless of which option you use, expect it to be slow ( compared to a similar Nvidia GPU ). I was running a 6800 and ended up just buying a 4060 ti 16gb and putting it in the PCIe x4 slot. It's way faster.

[lucid yoke]

Theres a problem with amd and ai lol

no cuda cores on amd and you have to emulate those

hence slow

[lime pivot]

Hello, guys can someone help me. there is always an error when I try to input the dataset in config file. Right now this is what I do:

Click "Prepare training data set", then " img > 14_x1y2 guy > x1y2 guy" is created

Then, in "Dataset config file" I input the file path:

"C:/Users/DESKTOP/Downloads/stable-diffusion-webui-master/kohyatraining/DarenTraining/img/14_x1y2 guy"

but there is an error

[lime pivot]

automatic 1111, sd 1.5. any help is greatly appreciated

[brittle lantern]

@fossil dew so I’ve added the file into wildcards and so when I go into wildcard manager I don’t have a wildcard file or the collection named what u told me to name it? Is that normal or do i need to do sum else?

[lime pivot]

for reference im follow this video: "https://www.youtube.com/watch?v=xXNr9mrdV7s&t=1083s" (from 17:06)

[fossil dew]

have you clicked refresh wildcards ?

[brittle lantern]

I have and it doesn’t show up

[fossil dew]

wth*

[brittle lantern]

Ohhhh

Ok

It has to be named

“Wildcards”

And I forgot the s

[fossil dew]

you got it?

[brittle lantern]

Again

💀

[fossil dew]

🤭 it's ok- i guess it does matter then, my bad

is it showing up now?

[brittle lantern]

Just transferring the txt

[fossil dew]

ok lmk

[brittle lantern]

Found it

It’s there

[fossil dew]

😄 hooray!

[brittle lantern]

So in promoting I just put “random anime clothes”?

[fossil dew]

__random anime clothes__

yes

[brittle lantern]

Ye thank u so I don’t specifically clothes I use that

Understood

[fossil dew]

yep 🙂 and you can add/remove/edit it however you need

just an example file for you ❤️ I hope it works out well for you!

Also you can put a name at the beginning with a : so you know what codes what 🙂 just a lil tip

[brittle lantern]

I understand

[lucid yoke]

scam

[north zodiac]

Hi all,

I managed to fix the issue I was having the other day with setting up ForgeUI on my AMD machine thanks to a chap on Reddit. However, I'm now experiencing another issue when trying to generate an image.

RuntimeError: CUDA error: operation not supported 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.

Specs:
AMD Radeon RX 6750 XT (ZLuda)
Pytorch: 2.3.1
RocM 5.7

Let me know if any further info is required 🙂

[slow cedar]

Ohhhh we're talking wildcards

Love wildcards

[brittle lantern]

yes can u send me the ones u had

rem when we tried doing it when we first me but didnt work

[fallow knoll]

its probably the same zip file i was gonna send you iirc lol

[silk bronze]

If I recall correctly I had to go edit Python source code for one of the Torch dependencies to get it to work - I was running ComfyUI though, not sure if forge is different in that regard.

Does it give you any specific information about where the error is being thrown?

[lime pivot]

Hey guys! Sorry to ask out of nowhere, but I have a favor.

I’ve been trying to make a LoRA for like 2 months now, and I’ve learned a lot, but I just can’t get any results. Right now, another person is helping me and said the reason I can't make a LoRA might be due to captioning problems. We’ve been going back and forth on many possible causes for my lack of success, not just captioning, but so far nothing has worked. Tbh, I’m just feeling a bit demotivated right now because I’ve put in a lot of effort and still haven’t seen results.

So my favor is: could you let me borrow a set of training images + captions that you’ve successfully used to make a LoRA?

That way, I can figure out whether the problem lies in my training images or captions, my parameter settings, or something else entirely. If I use your images and still can’t get a working LoRA, then I can safely assume it’s not the dataset and can cross one more thing off the list of possible causes.

I just want a taste of success. I feel like I’ve been heading nowhere despite trying hard, and I want to know what it feels like to successfully create a LoRA at least once, even if it’s using someone else’s training images and captions.

Would anyone be able to help me?

(sd 1.5, automatic 1111)

[north zodiac]

I'm not sure. It's all a bit vague to me. I have copied the CMD log though if this helps?

[slow cedar]

I can but like ret said he probably will send you the same ones I have. I think I sent him mine at one point

I make monster prompts with wildcards

Wildcards are my thing 😅

Scam

Don't click

@errant locust leave broski, your scams ain't gonna fly here

Shoo shoo scammer

[fallow knoll]

you should see the collection ive built XD

[slow cedar]

Dawg you gotta show me

I've been out of the game for a bit

😂😭

[fallow knoll]

@slow cedar

[slow cedar]

Oh my gooooddd

Can you send me those?

[fallow knoll]

inside every folder is 4-8 wildcard files, inside every file is 20-30 lines lol. all of them are thematic in some way. theres 84 folders in total

[slow cedar]

I need context for 'honeytrap_protocol' 💀💀💀

[slow cedar]

Jesus

[fallow knoll]

spy related seduction, thats about it

[slow cedar]

You went on a wildcards rampage while I was gone Lmaaaooo

[fallow knoll]

according to chatgpt, even if generated 1billion images per second, it would take longer than the universe has theoretically existed before i ever saw a true duplicate image

[slow cedar]

That's insane man

Yeah slide those my way

I'll be using those tonight lmao

[fallow knoll]

lol i should warn you...its not for the feint of heart. and its quite a bit more random than what your probably used to. theres a mix of nsfw, body horror, emotional shit, abstract concepts etc etc

[slow cedar]

Not a problem trust

[fallow knoll]

we will see. lol

[slow cedar]

You're gonna see me saying the weirdest shit is 'super cool'

'oh my God this super gorey image is so pretty' 😂

[brittle lantern]

I got lots of bro I thank him and u

[tired grove]

ohh great, thanks

[gusty tide]

Anyone know how to fix it from doing this why isnt generating the sheet the way I put it

[slow cedar]

🙏

[neon orchid]

hello guys

my friend need a help :/

He followed all the steps but when installing stable diffusion, this error appears.

[neon orchid]

my friend have a AMD

no nvidia

and try to install Zluda

[ornate elk]

This setup is from the nvidia guide.
The Zluda set up would have a different name

[pale garden]

I am so so tired of stable diffusion freezing my pc. Why cant i find an amd guide anywhere

[ornate elk]

Have you checked my Zluda setups guides? (First link of the pinned messages)

Which GPU and webui do you use? And when does it freeze?

[timber oak]

hello there, does anyone know a bit about stability matrix?

My search function isnt working anymore

[pale garden]

I have searched for them but none of them seem like relevant information, just how to set it up with rocm. I use 7800xt with ryzen 9900. I used to use forge but switched to SD.Next because the CMD in forge recommended using SD.Next for AMD. Plus forge had a weird VRAM allocation issue. I always swapping things around and demanding more VRAM model space just to change parameters when the setting was changed for more allocation etc... Sd.Next seems faster but it has the same dang freezing problems with a 10,000 settings with no guide explaining anything

Honestly generations are all over the place in general. Sometimes its generating like a 4090 and sometimes like a GT 710

[lucid yoke]

Your trying to generate too much at once, theres simply not enough pixels for it to make a coherent picture

I recommend generating the images seperate and combining them later

[digital frost]

Hi i wanted to ask if this is possible:

I want to make lora that would generate real life things in minecraft texutre style so i can paste it to game.. is this even possible? where can i get the dataset? how do i get data set? how to make all this work? sorry but i'm really new into Stable Diffusion and i don't really know how and where to begin

[cobalt moat]

Hello, I'm building OpenInteractions - An Open Source Global Infrastructure for Real Time AI Voice and Vision Interactions.

Github: https://github.com/OpenInteractions/OpenInteractions

We'll be live soon tomorrow on Product Hunt

[ornate elk]

Oh so you search for a usage guide?

If you use zluda as Backend feel free to share the settings you used when it freezes.
I can't help with SDnext and wouldn't recommend using it over forge or Auto1111 with Zluda.
I used a 6700xt in the past without issues on Auto1111 Zluda.
And now I'm using a 7900xtx and no issues at all on Auto1111 / Forge or Reforge with Zluda.

[ornate elk]

First thing good to know is to not use models (checkpoints) that are larger than your available vram.

Also when upscaling dont set to much Hires Steps.
If available enable Tiled VAE, that reduces freezes.

Also disable/exit Wallpaper Engine if used. That causes freezes and even PC crashes when used together with SD.

[lusty pewter]

Just installed Forge and a few models, dont really know what im doing, i keep getting this error

TypeError: 'NoneType' object is not iterable

bunch of elitists on reddit are just throwing around big words and saying python a lot and I need someone to ELi5 or help walk me through possible solutions or how to troubleshoot this issue. I dont have any coding experience but ive been using lots of AI image generators for a long time

so what is this channel for then?

[somber loom]

Hey guys I help people with growing your YouTube channel

[fallow knoll]

you post here for help, but they will likely ask for a screenshot of your terminal log to see exactly what failed. takes a few minutes for someone to get to you, people have lives and stuff. but this channel is for asking for debug help. me personally, im dumb so i cant help lol

[lusty pewter]

wait, so this rodriguez guy just sent me to another server telling me to open a ticket, yet he's new here as well

[fallow knoll]

that was spam, dont follow that guy

[lusty pewter]

I can smell the curry from this chat, please ban this man

[fallow knoll]

CS1o or someone else that has a light green name like his are the ones you wanna listen to, when it comes to debug advice, id ignore any other links or server invites you get lol

also yeah, that is VERY clearly a scam

[lusty pewter]

said some choice words and was banned from the chat, lol, that directs me back to my original question, can someone help me with this, what do i need to provide in terms of error logs etc

[fallow knoll]

you'll need patience. @ornate elk whenever you have a moment king, thank you.

[lusty pewter]

I assume im in for a big learning curve here but Im excited. I hate chatgpt and midjourney etc all telling me what i can and cant generate haha

[fallow knoll]

lol the learning curve isnt terrible. the hardest part is finding models you like and then finding out how to prompt them to get good results. theres a lot of resources out there that talk about the different base models, but people post pruned or trained checkpoints/models on civitai and huggingface, so you can usually find whatever your looking for between the two of them.

the settings and parameters may seem overwhelming at first, but you can always look and see what settings the creator suggests for a model and just use that. usually yields decent results

[lusty pewter]

i have downloaded a few models from civit, i will go back and check what settings have to be what to make things work. i have a more than capabale PC and i feel like im decent at prompting, its just all the extra technical stuff that SD opens you up to

[sudden pike]

@pure tusk

[fallow knoll]

yeah. i was lucky, i just shared my errors with chatgpt and it helped me debug lmao but i didnt run into many issues setting mine up

[lusty pewter]

okay, forgive me if this is the wrong term, what is the best "client" to use. i downloaded forge but my buddy is in discord with me right now using comfy and not having issues, is there a "best" one to use thats less prone to bugs

[sudden pike]

i use AMD and its just stressful xD

[fallow knoll]

i use forge, but i started with A1111. Comfy gives you more granular control because of its workflows, but has a steeper learning curve. If you have an open mind and no experience or habits with another client, id recommend learning comfy. But its not without its issues and debugging lol ive got a lot of friends that use it and they are constantly complaining about an update breaking workflows etc.

[brittle lantern]

question what has higher qual on flux GGUF Q8 or FP8?

I rather wait here bud

[mystic dirge]

I cant for the life of me get Forge to not give me pixelated images

Also, the images get worse the more Loras I use

[fallow knoll]

well forge and A1111 are basically the same, Forge is just hyper optimized for speed and to reduce strain on your hardware. It has no impact on whether your output is high quality, that sounds like either an issue with your parameters or your prompts/model.

as for the images getting worse the more loras you use, well yeah, thats what happens when you try to cram a bunch of different mini model influences into the iteration of a model that may or may not play well with loras. I never use more than 2 or 3 loras at a time, and i avoid conflicting style loras or character loras etc.

[mystic dirge]

That makes sense, I'm using one for hands as well as other stuff and I suppose I could run the image through the program again and just inpaint better hands.

[fallow knoll]

yeah, i rarely bother fixing hands, i just press generate til it gets it right. but you can get embeddings that will do that, you dont need a lora to do it

[mystic dirge]

Embeddings?

[fallow knoll]

look up "lazy embeddings" on civitai. its what i use, they have one specifically for hands

you put it in the negative prompt, like you would a lora. but unlike a lora, its even more lightweight and wont slow down your generation time or impact the image style or the model your using.

it doesnt guarantee success, but more often than not my hands are fine since using it

[mystic dirge]

Im using flux1-dev and with some of my generations I would get weird fuzzy blobs in the final image whenever cranking up the CFG Scale (Not the Distilled one). SO It made it to where I thought I had to avoid even enabling negative prompts

Im new to this, so Im sure theres something Im just straight up overlooking

[fallow knoll]

ah flux, im not familiar with it, but it sounds like a VAE issue. though i could be wrong. someone will surely come along that is familiar with flux and help you out though

[mystic dirge]

Hmm, yeah I have noting there at all. Ill look around in the meantime

[fallow knoll]

unless the model says specifically that the VAE is baked in, you always need a VAE

otherwise youll get washed out, muted, greyscale or spotty images

flux could be different, but thats usually the case with every other type of model

[mystic dirge]

Yep, just looked it up, I should have one. I see three results on Civitai. Ill test it out

[fallow knoll]

best of luck friend!

[mystic dirge]

Hey man, you've been a great help. Ive been fighting with Forge for the past 2 days

[fallow knoll]

lol i aint the smartest, but i know a little.

[brittle lantern]

with flux like that u need vae, clip l, t5 encoder

[fallow knoll]

👆

[brittle lantern]

currently im messing around and im using this

[mystic dirge]

wow, ok Ill look for these. I thought I only needed one

[brittle lantern]

u only need t5, clip l and t5xxl fp16

[brittle lantern]

a finetune like rayflux v3 as it aio

u load it in like

a normal model

most flux models are not all in one

so u will need to add them sep

like base flux model

@fallow knoll lazy hand embed goes into neg prompt correct

[fallow knoll]

correct

as well as lazyneg

and lazynsfw if you dont want nsfw images

[brittle lantern]

dw im fine without that xd

[fallow knoll]

then you can slap it in the positive prompt and get even more chances at nsfw stuff, even when you didnt explicitly prompt for it lol its like gambling with boobs!

[mystic dirge]

ditto lmao

[brittle lantern]

my typa shi

ayo

i didnt even type things and got nsfw defo my typa thing

💀

[fallow knoll]

lmao

[mystic dirge]

Probably a dumb question, but what is t5? I have these. I assume that the t5xxl fp32 that I found is just an improved version of t5xxl fp16.

[fallow knoll]

t5 is the text encoder, its what allows flux to understand your prompts. VAE basically tells it to render the image without the usual imperfections you were having (i dont really understand the why of it or the how, but its necessary with EVERY model to have one), the CLIP is also needed for it to properly understand and render images from text prompts iirc

with other models, its already baked into the code of the model

[mystic dirge]

Ok, Im just wondering do I need t5xxl and t5? I found "t5 (GGUF)" on CivitAI and im pretty sure thats the file I need but Im not sure. The File name is different than the one @brittle lantern has on their setup.

[brittle lantern]

ill say flan t5 takes quite the gpu power for me

what gpu u got?

[mystic dirge]

NVIDIA 4090 24GB

[brittle lantern]

yh u will be fine then

i have RTX 6000 for what im using now

[mystic dirge]

So t5 GGUF is the right tool?

[mystic dirge]

My images are coming out in a cartoonish style now

[brittle lantern]

i mean i would use normal not gguf

[mystic dirge]

Ok, this is what I'm working with now

[brittle lantern]

and vae

and set the swap to shared always

[mystic dirge]

Ok, I found the VAE on HuggingFace

[mystic dirge]

I downloaded an Upscaler and put it in the ESRGAN folder. but it isn't showing up in the HiRes Dropdown menu

nvm. fixed somehow lol

[fossil dew]

How do you fix quality in forge without using high res fix?

[pale garden]

Even better would be knowing the kind of Forge settings YOU are using

[ornate elk]

Depends on what I want to generate ^^
And which model I use and if I enable hires fix or not.
Do you need the good hires fix settings?

Or in general something with sdxl/pony/illus model

[subtle breach]

Morning, people.

I'm tryna install stable diffusion with A1111 but I've been encountering a problem

Anyone mind helping?

[lucid yoke]

Whats the problem your encountering? And what gpu do you have

[subtle breach]

just erm give me ten mins please

i'm a bit busy

[lucid yoke]

Feel like thats the more important question since i dont know whats happening lmao

[subtle breach]

I have 8GB ddr3 ram

[lucid yoke]

Ahh no gpu?

If so i think i already know the problem

[subtle breach]

umm

go on

what is it?

[lucid yoke]

With only 8gb ram and no gpu (aka no vram) its trying to run everything on the cpu which is tragically slow

Stable diffusion, unfortunately requires a decent GPU (with a preference for Nvidia)

[pale ibex]

Has anyone tried running stable diffusion a1111 on a macbook air m2?

[subtle breach]

I see

[lucid yoke]

Hmm i know @brittle lantern uses a Mac for sd in the past

[subtle breach]

how about I send the link to the yt video I'm following for instructions?

[lucid yoke]

Well if you have no gpu and 8gb ram i can honestly recommend to not install it and instead look at web based image generation

Like midjourney

[subtle breach]

I just want something good and free

[pale ibex]

A simple prompt on my air m2 takes upwards of 50 mins to an hour, just wanted to know if those are normal speeds cause I expected it to be slower than an nvidia running pc but not THAT slow

that is without any loras

[fallow knoll]

dont we all lol

[lucid yoke]

Don't we all, I'm afraid your pc is not equipped for stable diffusion

[subtle breach]

aw shucks

well thanks anyways then

[lucid yoke]

No problem buddy, you can try looking on civitai to mess with it, you get some free credits daily

[subtle breach]

eh thanks

[lucid yoke]

Its not much but with sdxl you can make quite a few images before it runs out and you get 25 free a day

You get the same results but eh a little less freedom

[brittle lantern]

i use stability matrix easiest way for me

[vocal burrow]

Cf the pinned guide. There should be something for Mac in there.

Viking's stuff

[frail grove]

i just upgraded to 5060 ti from 3060 ti
and i get this error
what do i do

[pale ibex]

will check it out, thanks

[pale ibex]

I see, what are the speeds tho?

[vocal burrow]

Cause you re using your previous torch install which was compatible with your 3060 but not your 5060

Not sure what software that is. But if you spot a venv folder try renaming it, then reload the software. It should rebuild the venv by itself afterwards.

(if it's working then delete the old / renamed venv folder)

[frail grove]

too late im already reinstalling whole ui

xdd

but thank you

[brittle lantern]

On m2 pro MacBook Pro with 16gb ram on SDXL I get like 1it per like 3 ish seconds not great for flux I honestly deadass say skip any UI on mac and download draw things

If u have normal M2 will be slower

Unless u use Adetailer or the traditional Hi-res I just say download draw things on App Store and call it a day

[pale garden]

Im mostly thinking of settings that wont be too slow and also wont freeze my pc. You know all the optimal settings ranging from oppx runtime, the integrated settings, optimization tab etc. I am still getting the hang of generations in general but hires.fix setting cant hurt.

[pale ibex]

yeah just tried it on my m2 air and it took ten mins for a simple prompt, and I didn't even get the img in the end 😭

I'm considering just building a new pc but gpu prices are not favorable rn so not sure what to do

[brittle lantern]

if ur in NA ur cooked

for prices

[upper linden]

been getting this error every time i try to make a image, GPU: 7800xt, Drivers: 25.4.1

RuntimeError: mixed dtype (CPU): expect parameter to have scalar type of Float

[upper linden]

something is forcing to use to my cpu i assume

[timber oak]

hey there

comfyui is complaining loudly to me that i am missing UnetLoaderGGUFDisTorchMultiGPU

however i have already installed all custom missing nodes

and updated to the latest version

i have downloaded it several times

switched the version

try update

but it still appears under missing

i have uninstalled and reinstalled it

it still shows up in missing nodes

[restive thorn]

Hello, I'm using SD directml (WebUI) on a Windows 11 system with RX6800xt (AMD) gpu. It has a lot Vram but it seems like it isn't used correctly and I have low technical capabilities inside SD as a result. Tried switching to onnx+olive-directml on SD-next but have gotten a lot of errors and didn't manage to fix all of them. If anybody is willing to help and has some experience with amd gpus, could you help with setting up an SD version which will work relatively smoothly? Because as it stands, I'm only capable of using up to 3gb of my vram, even though all 16gb are reserved by SD-directml version. 🙏

[lucid yoke]

I recommend looking into one of the guides of CS1o in the pinned messages, directML is not recommended as it uses a LOT more vram then zluda does (if your gpu supports it)

With amd i recommend forge ( its what most people with amd use here)

If you get stuck i recommend posting here again since cs1o is fairly active here

[vocal burrow]

I saw your previous message. There is no "suport desk discord" or anything alike.

Don't click random stuff from random people dm ing you.

[restive thorn]

Yeah, I already figured, it was really dumb of me to even think that considering the separate invite link🗿

[vocal burrow]

You figured it out so we re good

[ornate elk]

Delete the venv folder and open up a cmd and type
Pip cache purge
Then rerun the webui-user.bat

[restive thorn]

Oh, hello there! I was just about to ask for help with a problem I have!

I'm having problems while trying to install Forge-Webui with ZLUDA
First issue I encountered: output was something like "torch wasn't compiled with CUDA enabled"
what I did: Install PyTorch with CUDA 12.1 support

• pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
  (while in SD's venv)

Current issue: ran into RuntimeError: CUDA unknown error - this may be due to an incorrectly set up environment, e.g. changing env variable CUDA_VISIBLE_DEVICES after program start. Setting the available devices to be zero.
Deactivating virtual environment...

I have adrenalin 25.1.1., downloaded zluda from the guide, rocm 6.2

actually sidenote: I have integrated graphics on my cpu, but I disabled it in the device manager
My main gpu is rx6800xt

[upper linden]

I did that still same error

[ornate elk]

Try update the webui by opening up a cmd inside the webui folder and run git pull

Or by adding git pull into the webui-user.bat

Then delete the venv folder and the .zluda folder. Boath are in the webui folder

[ornate elk]

Hey, can you copy and share the full cmd log when launching the webui-user.bat?

[restive thorn]

Put it into a txt file because it's too long to send as a message

[ornate elk]

Can you try to generate a normal image?

[ornate elk]

Hmm looks correct, strange that you get that error

Try delete the venv and the .zluda folder (both are in the same main folder) and then relaunch the webui-user.bat

[restive thorn]

Did that. Same error, appended the log

[sudden pike]

question. Normal that i can only use a batch size of 1 on OneTrainer

it works for me... but uh yeah it uses hella amount of VRam xD

and OOM / Driver Crashes are coming fast

[sudden pike]

have this too sometimes

that's VRam issues

try to lower resolution for testing

also what Driver version do u have?

[restive thorn]

nah, I didn't generate anything yet, it's the startup log😁

[restive thorn]

Adrenalin 25.1.1

fekken AMD gpus bro🚬

[sudden pike]

i know the feeling

also an AMD user

[restive thorn]

my condolences

[sudden pike]

AMD is just bad with AI stuff

[restive thorn]

let's take it to general chat shall we?😅

[sudden pike]

if ya want?

[restive thorn]

yeah just don't wanna trash the tech support chat

[upper linden]

yeah

[restive thorn]

I'll try disabling the integrated graphics in bios too
Considering that it might be a vram issue and it doesn't have a lot of it, maybe SD is choosing it instead of rx6800xt at some step
Update: didn't help

[upper linden]

i try and it still give me the same error

[restive thorn]

I'm finna sell a kidney and buy the 4090 at this point

[ornate elk]

@upper linden @restive thorn Sry guy's I'm still on vacation so can't help further.
I'll be back on Friday/Saturday. Feel free to DM me so I dont forget it.

[restive thorn]

Ohhh that's ok sorry for bothering you then, have a good one!
I'll dm you my case

[upper linden]

Ok

[rotund portal]

I don't know if it's a recent update or something changed on my PC but SD-Forge has been weirdly slow for me. I'm running a 4090 and it seems to hang for a bit on weird actions. E.g. Auto Detect from IMG Size causes a hang?. Changing from txt2img leads to a whole slowdown of pc until it loads.

[pale garden]

You have the same gpu as me. Try to see if you can generate outside of fp16 and fp32. If that doesnt work try to add these commands --no-half --no-half-vae --precision full

[ornate elk]

For that I would need to know which settings do freeze your PC xD

Also never use --precision full --no-half
Its a massive slowdown

For Forge on AMD use --use-zluda --cuda-stream --attention-quad

[ornate elk]

Make sure you dont have Wallpaper engine running in the Background also add --cuda-stream --cuda-malloc to the webui-user.bat

[rotund portal]

This makes things a lot better but I still have weird behavior with switching to inpaint and trying to auto resize or use png info. It pretty much just freezes Firefox. Like tabs will just load then.

[ornate elk]

Whitelist the webui in any browser adblocker like ublock origin

Adblockers can cause issues in the img2img tab for example

[rotund portal]

That did it. Wow... I'm so annoyed it was that. I spent all day reinstalling it...

Just a question. Why does Wallpaper Engine need to be turned off?

[ornate elk]

Because it causes slowdowns and freezes and even can crash the PC when run at the same time with SD

For the best performance just exit/close (not freeze) it

[rotund portal]

Yep. Okay. I've done this. I should also lower my GPU weight right? Since Forge seems to want to max it almost.

I'm using SDXL

[ornate elk]

Use the default weight value forge sets (not Max, but near)

Your on a 4090 so there shouldn't be any problems with sdxl based stuff

[rotund portal]

You'd think. Image generation is at the same speed. It's just nav stuff that's broken. So like applying styles etc.

Might try with Chrome and see if that fixes it.

[ornate elk]

I dont have any problems with Auto or forge using only Firefox with ublock origin whitelisted

So maybe its an other extension that breaks yours

[rotund portal]

I'll try disabling extensions. That could be it too.

[ornate elk]

Also dont use multiple adblockers or script blockers

Or AV stuff

[rotund portal]

I just keep to UBlock Origin

[ornate elk]

Thats the best and only trustworthy one

[pale garden]

I dont think it would hurt to do a little process of elimination. Also he could try Launching Web UI with arguments: --use-zluda --cuda-stream --attention-quad --opt-sdp-no-mem

I think opt-sdp-no-mem is needed for 16gb vram???

[ornate elk]

Opt-sdp-no-mem is not needed and can't be used together with --attention-quad

Let me know your settings when you get the freezing issue again, then I can check what's causing it

[mystic dirge]

Im using RayFlux, and my cmd window is giving me this error message when generating. Can anyone help me figure out what in the Forge Webui is causing it?

Error running process_before_every_sampling: P:\Stable Diffusion Forge\webui\extensions-builtin\sd_forge_controlnet\scripts\controlnet.py
Traceback (most recent call last):
File "P:\Stable Diffusion Forge\webui\modules\scripts.py", line 892, in process_before_every_sampling
script.process_before_every_sampling(p, *script_args, **kwargs)
File "P:\Stable Diffusion Forge\system\python\lib\site-packages\torch\utils_contextlib.py", line 115, in decorate_context
return func(*args, **kwargs)
File "P:\Stable Diffusion Forge\webui\extensions-builtin\sd_forge_controlnet\scripts\controlnet.py", line 561, in process_before_every_sampling
self.process_unit_before_every_sampling(p, unit, self.current_params[i], *args, **kwargs)
KeyError: 0

[brittle lantern]

HiDream Full or Dev can run on 24gb vram Nvidia GPU right?

[mystic dirge]

Ive never used those, but the newer GPUs seem pretty capable so I'd be surprised if they couldn't.

[brittle lantern]

I’m running a slightly older RTX 6000 24gb

I can run all models

Pretty much fine

Never tried HD since I know it’s the most taxing model currently on open source

[mystic dirge]

Ill give it a go. RayFlux is treating me pretty well right now.

have you ever gotten that error message before though? It still generates and image for me, but the error seems kinda odd

[brittle lantern]

No

I just use AIO set the thing from CPU to shared and use 2m and SGM uniform at 30 steps

And have fun with it

Can’t help u with that error unfortunately

I’m going between V3 and photo real

[mystic dirge]

What are the full names of those two models? Would you say they are better than flux? I'm trying to generate realistic looking people

[brittle lantern]

On RayFlux model page

It’s literally V3 AIO and photo real or photo plus

https://civitai.com/models/1056935/rayflux

[mystic dirge]

oh, damn lol I see the tabs now. v3.0 AIO and PhotoPlus. My bad

I thought it was another model entirely

[brittle lantern]

Nah XD

[mystic dirge]

And you dont need refiners for these? Im using an upscaler but thats it (and even then Im getting the "4x-UltraSharpV2.pth' is not a 'ESRGAN' model" warning).
But with some generations I get weird warpy backgrounds and Im trying to stop that

[brittle lantern]

Yeah that’s fine it will auto switch to DAT dw about that

I use the V2

[mystic dirge]

Ok, nice

Thanks for the tips! 😄

[brittle lantern]

This isn’t RayFlux but it’s with adetailer and hi-res 20 steps 1.5x at 0.4 denoising

[mystic dirge]

Thats a really good image

[brittle lantern]

Cheers it’s ultra real fine tune

[mystic dirge]

Is the Base model just flux1-dev?

[brittle lantern]

No

[brittle lantern]

Here I mentioned it

[mystic dirge]

Ah, I didnt know that was the base, I thought it was an extra

[brittle lantern]

https://civitai.com/models/978314/ultrareal-fine-tune

[mystic dirge]

Thankyou so much

[brittle lantern]

Got u

It’s a manual one so u have to add VAE, Clip L and T5

RayFlux 3 u just load it in and run it like a SDXL model

[mystic dirge]

Oh really? Ive been running all of these with RayFlux regardless

[brittle lantern]

With RayFlux 3 u don’t need them

With V1 and photo real u need those

And ultra real fine tune

Also u using 4090 use 2200

*22000

[buoyant lava]

Guys how do I make hands not bad

[mystic dirge]

lol back to the drawing board I go 😅

[mystic dirge]

Haven't tried it myself yet, but Ive heard that some people run the image back though Inpaint with a LoRA trained on hands

[buoyant lava]

can you suggest a good hand lora for like anime style

[mystic dirge]

Ive only been doing realistic stuff so far. CivitAI is where I get the majority of my models though. I judge them by the pictures and the amount of downloads before I try them out.

[buoyant lava]

Can you help me find a hand lora

[mystic dirge]

sure, gimme one sec

Oh, what is your base model?

[sudden pike]

you can try negative prompts

[buoyant lava]

sd 1.5 i think

[buoyant lava]

tried, either keeps the hand mangled or just makes it dissappear

[mystic dirge]

@buoyant lava This one seems pretty well reccomended.
https://civitai.com/models/200255?modelVersionId=486156
It also seems to work for realistic stuff too. i downloaded it the other day but I havent really used it yet. The base model KitKat recommended seems to handle realistic hands pretty well straight out of the box

[sudden pike]

Do you use an LoRa?

or is it a self made prompt and character?

[buoyant lava]

im trying to fix the hands using inpaint

[buoyant lava]

i read the comments and many are saying that it sucks

images dont look taht great either

a bunch are still really mangled

[mystic dirge]

really? one sec. I downloaded the flux version. Like I said I was basing it off the amount of downloads

Ill look deeper for another

[buoyant lava]

i think i found one with good rating

i will try it first

https://civitai.com/models/278497/hand-fine-tuning?tags=4

[mystic dirge]

The comments on this one look better. Let me know how it goes. I would have overlooked this one because the images look a little hazy even on the fixes. But in their defense it could just be the base model or prompt style.

[buoyant lava]

Ok, the hand looks better now, no longer looks like claws, still a few artifacts

[sudden pike]

you can use multiple LoRa 😄

[buoyant lava]

And now my hand randomly turned into a foot

So cursed

[sudden pike]

LMAO

i wanna see that

[sudden pike]

xD

[buoyant lava]

it was doing fine until i prompted “thumb out”

heres original

i want something like this pose but idk how 2 prompt dor that

[sudden pike]

what model do u use?

[buoyant lava]

anything v5

should i swap to meinamix or anylora

[sudden pike]

what base model?

[buoyant lava]

SD 1.5

[sudden pike]

ah

[buoyant lava]

@sudden pike Are you good at prompting

[sudden pike]

[buoyant lava]

I’m trying to use depth hand refiner

It still doesn’t work great

Will upscaling help?

[ornate elk]

Your on 4gb vram, so the best you can do is to use Forge Webui. With that you even can use sdxl/Illustrious models at an okayish speed. But it also boosts 1.5 gen time.
If you want to have good hands on 1.5 models the key is to use hires fix as this can fix hands and faces

[buoyant lava]

Settings for hres?

[ornate elk]

Upscale by: 1.5
Hires steps: 10
Denois: 0.5
Upscaler: Resrgan4x+ Anime6b

And if your on forge always enable
"Never oom for Tiled VAE only" when upscale

[buoyant lava]

Tried upscaling but hands still bad

[buoyant lava]

I’m on auto1111

[ornate elk]

Then install the Tiled Diffusion extension for its Tiled VAE setting and Adetailer for face fix

[buoyant lava]

I’m trying to fix hands

[ornate elk]

Also make sure you have --xformers --medvram in the webui-user.bat

[buoyant lava]

And what does tiled vae do

[ornate elk]

For the best performance

[ornate elk]

It speeds up the VAE process when upscaling

Reducing time and freezes

[buoyant lava]

Upscaling for me is already instant

~10 seconds at most

[ornate elk]

Ah okay can you show me an example for before and after Upscaling?

Upscaling normaly takes longer

[buoyant lava]

Uhh

Well I do notice a marked increase in resolution

[ornate elk]

If you have no issue then you can also try upscale by 2

Also how long does a 768x1024 sdxl image with 30 steps, Euler a takes for you?

Afk

[sullen nymph]

Is there a way to stop SDNext* from being unloading the LoRa network just to load it again when using detailer?
I was having a similar issue before when using 2 LoRa but increasingLoRa Memory Cache fixed it, but the issue persists with Detailers

Image example of the behaviour below.
20s Image, 7s Detailer, 24s Network (Unload & Load all LoRa twice)

[ornate elk]

Is that really forge webui your using?
For me it looks like SDnext

[sullen nymph]

Yes SDNext, my bad

[clear ridge]

Hey people
Can anyone here help me with runpod, trying to create a Lora. Walking into too many problems. Especially when I do all the premilary things, press start training and nothing happens. Endless clicking

[vast crest]

Hi guys does anyone know how to use this version of Colab: https://colab.research.google.com/github/Jelosus2/Lora_Easy_Training_Colab

[sullen nymph]

Any info on a way to improve or completely fix it? just increasing the LoRa Memory Cache doesn't fix that

[sudden hazel]

Why can't i see "create mask area"
I want to paint the regions

[sullen nymph]

that is a sketchy af group

looks like a place for scams

@ebon viper from where is it sourced then

deleting messages, nice

[vocal burrow]

yup it was.

there is no "official support/ticket discord" or whatever.

L O L

saw you @pallid osprey , please let me know about your next scam account.

[lucid yoke]

@pure tusk sorry to bother but: UserID: 1376996405565984799 @ebon viper scam account of the "helpdesks" linking and @pallid osprey UserID: 1379141584305524817
its happening a lot lately

[lime pivot]

Hey guys, apparently fkunn1326/openpose-editor has been dysfunctional for some time now. The tab won't show even if you install it, and I've seen people with similar problems. I've also tried downloading the sd-webui-openpose-editor in the Extensions > Available tab, but I don't know if it didn't work or I can't find it. So, are there alternatives for SD 1.5 AUTOMATIC1111?

I also tried to open Latent Couple, but I'm having the same problem.

[tawny saddle]

returns from the abyss

anyone ever heard of pifuhd

anyone got that trained model

i cant seem to shake hands with the server

ssl certificate

-screams-

[tawny saddle]

nu sadly

oki

a oof

thought u was a real human for a sec

this wut i call bot infestation

the inventation user isnt even ur own, that you sent to me xd

kevin 2.0

whole server prolly full of kevins

xd

[sudden hazel]

i have a model for ADetailer, where do i have to put it, so i am able to choose it in SD?
I'm on Automatic1111

[tawny saddle]

for the record i didnt paste it into browser either

xd

[tawny saddle]

[fallow knoll]

Another one for the block list meat grinder

[tawny saddle]

xd

[fallow knoll]

say what you want about mancho man, but hes dedicated to his craft

thats rare these days

[vocal burrow]

not dedicated enough to craft a proper script.

[vocal burrow]

that one ? https://dl.fbaipublicfiles.com/pifuhd/checkpoints/pifuhd.pt

it downloads just fine on my end

[tawny saddle]

would chu mind tossing my way o^o

[clear ridge]

I keep getting this error when trying to upload, I’ve opened up 3 sd templates same thing

[vocal burrow]

dont

good bye scammer

[vocal burrow]

if you joined the discord from the deleted message, I invite you to leave it. it s a scammer discord.

There is no "official support discord" or whatever.

[clear ridge]

I see

Ok cool

Not a scammer :/

[vocal burrow]

what you mean ?

[vocal burrow]

Also, this screenshot give basically zero information. we don t know what you re trying to do, what software you re using, etc

[clear ridge]

I thought this discord group was for support using stable diffusion

[vocal burrow]

no, it s just some scammer group trying to get you to install ransomware.

Any official support is done either via email or stability.ai official website directly.

This very channel is community driven.

There s no point splitting the community into multiple discord. It would make everything harder to look for, people would not be able to learn from others by searching the past messages, etc

[sullen nymph]

I like the very large group with 0 bot moderation for people posting links

[formal rapids]

How to vote to get the free lumens ?

[vocal burrow]

what do you mean ?

free lumens ?

[tawny saddle]

if u want free lumens

jus step outside when the sun is out

to get them lumens mate

buh dum tiss

[sudden hazel]

i just switched over from forge to A1111 and did my first generation... damn it runs VERY slow.
RAM usage is 100%
GPU too of course

is this normal with the first generations?
wanted to generate 4 Batches, each in 5 Steps

[ruby yoke]

switch to comfy

[tawny saddle]

wha

how would using a diff Ui have anything to do wit slowness xd

isnt that just an Ui, and have non to do wit how stable runs underneath

wouldnt that be like, switching ur keyboard out, to make ur pc faster xd

if anything id say comfy Ui would make using stable slower, with all the spaghetti and like prolly 200+ options xd

ComfyUi looks like, to me, the inside of a breaker panel of a skycraper

comfyUi be like

xdd

a111 be like

[hidden furnace]

I need help, I'm getting this "RuntimeError: mixed dtype (CPU): expect parameter to have scalar type of Float" and can't generate any image, I did nothing different, didn't use any new model, lora, didn't change settings, I was just testing prompts, checking the results, the only different thing I did was to change the sampling from euler A to DPM++ 2M SDE, I left to take a shower while the images were generating, and when I came back I had this error, I tried restarting windows, but I still can't generate nothing, I get the same error

[tawny saddle]

todd howard be like, "and again, it just works"

[hidden furnace]

the only different thing I noticed is that now, when starting up the auto1111, the attention optimization has changed from doggettx to InvokeAI

[lucid yoke]

A1111 is hella outdated and lacks support for many models, comfyUI processes things a faster and with a good workflow it can save actual seconds

When going from a1111 to SwarmUI (basically comfy backend) i improved with 10-20s depending on the model

[tawny saddle]

oop well, i wouldnt know, i downgraded to 1.7 bc 1.10 was slow and lacked responsiveness, also seemed to not understand wut i mean by orange beak whilst using the same model and prompts i been using xd

[lucid yoke]

Granted those were the earlydays, a1111 did get a few updates but it's not recommended unless your on amd

[tawny saddle]

and that would be me xd

AMD igpu

vega 7

i bet

im prolly the only person in the world still using 1.7.0

xd of a1111

[hidden furnace]

just reinstalled auto1111 and still the same error, I have no idea how to fix this

ok, I give up, this is so stupid, I can't understand how it simply stop working out of nowhere

[tawny saddle]

blop

ask them for help

[tawny saddle]

CS1o....the ppl need u mate, where are u...

[fallow knoll]

on vacation

[fallow knoll]

see?

[polar crest]

since, ive managed to manually install CLIP!

however, now it says i can't install "open_clip"

any ideas?

[tawny saddle]

xd

[tawny saddle]

oop

[ornate elk]

Open up a cmd and run
Pip cache purge.
Then delete the venv folder of the webui.
Then relaunch the webui-user.bat
If you get an error please copy and paste the Full cmd log here.
And dont install stuff manually it won't work.

[ornate elk]

In around 1 day I'm back!

[tawny saddle]

xd have a gud vacay

glad i made u laugh wit me CS1o aerial call

xd

[ornate elk]

"Commit hash none" means you either didnt git cloned the webui or you moved the folder afterwards.
I would recommend upgrading python to 3.10.11 64bit and then git clone the webui again.
Delete the old webui folder before.
For the best guide check the first link of the pinned messages.

[polar crest]

understood! I cant remember if i moved it, i THINK i just extracted it to a spot inside my users folder after download.

but anyhow - ill start from scratch and git clone the webui using the upgraded version of python!

Thanks so much for your help! I'll let you know if anythign else goes wrong!

ah

it was malwarebytes all along

was stopping torch from working, clip form downloading, open_clip from downloading

[ornate elk]

Ah yep AV programms can make problems

[polar crest]

absolutely, but i wouldnt have figured it out without your suggestion to reinstall with a new python version. For some reason malwarebytes didnt send error messages when i re-installed the same version of python over and over and re-tried the process....

but for some reason when i downloaded 3.10.11, it decided to re-show them - so this time i noticed them!!!

thanks again!!

[ornate elk]

No problem 🙂

[ornate elk]

Hey, what's your GPU and please provide a full cmd log

[livid stone]

Hello, is there a setting i can toggle for images to be generated directly as "upscaled version" -- in case i need to afk from the computer?

[ornate elk]

Hey, if you use Auto1111 or Forge or reforge then you can just enable hires fix

Set hires steps to 10
And denois to around 0.5 or 0.4

[livid stone]

i tried enabli9ng hires fix, but it wasnt quite the same finished image once i clicked "upscale" to the generated output

im using Auto1111

[candid valley]

Hi, everybody

[ornate elk]

You can define the output with the denois value.
Lower denois means it doesnt change to much 0.4 is a good start.
Don't go above 0.5

[candid valley]

I'm AI developer specialized in generative AI

[livid stone]

ill try messing with the options thanks

[candid valley]

and also llm, langchain and RAG

[buoyant lava]

HELP! I can’t find the T2I adapter model for color, for stable diffusion version 1.5

[tawny saddle]

a b c d e f g

[vocal burrow]

don't get on that discord (I cleaned the message)

it's a support scam trying to get your wallet / install ransomware

[opal crag]

uh

i got this msg when i chatted in this server

[vocal burrow]

in dm I guess ? they're trying to hide nowadays lol

[opal crag]

i was like since when do staff send dms

[vocal burrow]

very very very unlikely

and writing "super buff man" in my bio won't morph me into Schwarzenegger

[opal crag]

@ Cho_lli♥

[vocal burrow]

stability staff have the according discord roles to show for it.

[opal crag]

and they dont have time to dm everymemeber who asks for help XD

[vocal burrow]

"Community guides" are not part of stability staff, but it's a role given by staff to trustworthy / mod / etc people.

[opal crag]

btw are u good at stable diffusion cuz im day 1 and need some help in #🏞｜general-with-images

[vocal burrow]

Why do I get different results from civit.ai/my friend/alien lifeform examples :

1/ --xformers (and some other flags but mainly this one) depending of its version can make stable-diffusion non deterministic. Meaning that for the same prompt+settings running on different hardwares you'll get different results. The output should still correspond to whatever the prompt is.... It might just be slightly to very different if you're unlucky. So yeah, some command line arguments can alter the outputs of your prompts and your hardware might also do the difference.
2/ Do you have all the loras, embeddings mentioned in the prompt ?
3/ Are you really using the same settings ? Copy pasting manually what's inside the "Generation Data" boxes is not enough. Civitai is hiding a bunch of data there. If you want to get the exact same settings you'll have to click the "Copy Generation Data" button at the bottom of the image page, paste its content into Auto1111's prompt field and then click the blue arrow under generate to automatically apply each value to the correct field. Hidden values can be; clip_skip, ENSD, token merging ratio, etc
4/ Did you have any overrides set beforehand ? They should show up at the very bottom of the page (some will probably show up too if you use data from civitai like I mentioned in 3/)
5/ Maybe it's using some extensions that does not record its settings in the metadata
6/ Maybe you're using different versions of some extensions/auto1111/lora/models/etc that yield different results
7/ Are you using the same random number generator ? GPU / CPU / NV (Settings->Stable Diffusion->"Random number generator source")
8/ Maybe you're using a newer version than OP. One that includes a "seed breaking change", cf https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Seed-breaking-changes

old copy pasta of mine but it should still be relevant from what I see

without even clicking on your screenshots I can see that you're not using the same resolution so you'll get different output for sure.

[opal crag]

lemme max the res and see

[vocal burrow]

(and yes, avoid nsfw stuff in here indeed :p)

[opal crag]

ill turn on some other settings

[vocal burrow]

won't work

you're still using different settings

this is not a square output

[opal crag]

i mean i dont need the exact same thing, like the style if its close enough

[vocal burrow]

like stated in my copypasta there should be a "generation data" field somewhere next to the picture, just click on the copy paste button,
copy paste everything into the "prompt" field of your A1111 instance, press generate and it should apply everything for you.

[opal crag]

okok

[vocal burrow]

make sure you also have installed the loras mentioned in the prompt

[opal crag]

ohhh

[opal crag]

i copy pasted the generation data, idk if it had any extra loras

and the img ratio should be same

[vocal burrow]

Also, I would add a shortcut for the "clip_skip" value if I were you , add "CLIP_stop_at_last_layers" to the Settings -> User Interface -> Quicksettings list

[opal crag]

okay ill see

[vocal burrow]

and looks like there is no loras being used

and 129 is a ridiculous ammount of steps

for now, just try to recreate the picture. you can tweak it afterwards

[vocal burrow]

and click the "blue arrow" below generate to apply all the values into the field

cause right now you have a lot of garbage stuff in your prompt that should be applied instead

[opal crag]

oh figured😂 let me fix my pc crashed too ig cuz of samples

[opal crag]

idk how to do this

paste clip stop at last layers in the quick settings list?

ohk i did

[vocal burrow]

press the reload UI button afterwards for it to show up at the top of your UI

or just close and relaunch webui-user.bat

[opal crag]

oh nice

omg its close

yoo thanks man its so good

[vocal burrow]

you probably made a couple of change from the original civitai pic

hard to tell without seeing a screenshot of your webpage and a link to the civitai pic

that's why you still get different output (that or different xformers, seed breaking changes, etc like mentioned in my copypasta)

[opal crag]

its fine i didnt want exact anyway i just wanted to learn how this worked to make my own stuff

[tawny saddle]

ah geez

I luv 1.7.0 so much

it just works yk

[opal crag]

love u bro

[vocal burrow]

for information, you can hover on every setting in A1111 and it should give you a short explanation about what it does.

[opal crag]

okayy

what is embedding btw

some of the creations using them

do i just install them like i did with the model

[vocal burrow]

tldr, cfg scale "how close it should follow the prompt" it has a huge impact. So does clip_skip but it's a bit more complex to explain. and there's pratically no point going over 40 steps.

[vocal burrow]

embedding is kind of a generic word. In the context of stable diffusion you're probably thinking of "textual inversion".
To be short, they're "patch" that you apply to a model in order to add new concept to it.

embedding / textual inversion is a very primitive method though. Loras work similarly-ish and are better.

just remember that embedding/loras are trained for a specific type of model (sd1.5, sdxl, sd3, etc) and you can't use a sd1.5 lora with an sdxl model (or derivative).

[opal crag]

also ive seen some img2vid
can stable diffusion do that?

[vocal burrow]

there are many ways to do videos with stable diffusion. The old ways with "deforum" or "animatediff" for instance. And the newer ways with "SVD" or using non stable diffusion stufff like "wan" or "lxtv"

lots of terms for you to google. You'll probably have to install forge or reforce which are forks of automatic1111 / A1111 / stable-diffusion-webui or comfyui

[opal crag]

😞 too much for now ill learn slowly

[vocal burrow]

probably for the best indeed. Start slow sith txt2img

then loras

then stuff like controlnet, adetailer, etc

[lucid yoke]

If they want to do video, you absolutely do not use forge/a1111, swarmUI/comfyUI is the ui for that (tough while swarm is more beginner friendly, its best to switch to it once you got a bit comfortable with the terms)

[vocal burrow]

sure but they began like yesterday, so yes, baby steps.

[buoyant lava]

Got on what discord

[vocal burrow]

some scammer sent a discord invite before that.

the message has been deleted since then.

[buoyant lava]

Oh

Where do I get the colour adapter for T2I

And for SD 1.5

[vocal burrow]

https://huggingface.co/TencentARC/T2I-Adapter/tree/main/models ?

[buoyant lava]

That’s for sd 1.4

[vocal burrow]

that's all there is, there is a coadapter for sd1.5

[buoyant lava]

doesn’t work with auto1111

[snow eagle]

Hello everyone. We are trying out Stable Fast 3D for low-weight poly mesh but we keep hitting the file size error 😭

[tawny saddle]

whose....

We

imagine u ask tech support a question and they gloss over everything and the only thing they reply with is, "whose.. we"

xd

[potent sun]

I'm looking for a model that I can feed papers to and make it sort a series of prompts.

Prompt order is important for some models and if I have the paper with all the info about the model, I would like a LLM that can make this task easier.

[tawny saddle]

wha..

feed papers? ovO

like a fax machine? ovO xd

look nu further than DunderMifflin, Dunder Mifflin can fill all ur paper and paper product needs

xd

[lucid yoke]

assuming you want it to read documents and use that as a information source or you want to give it a document and have the contents as a prompt?

your being kinda vague

[potent sun]

I was thinking on giving it stable diffusion papers

and similar

[lucid yoke]

so you want it to make image prompts?

[potent sun]

no

[lucid yoke]

then you dont need to give it papers AT ALL

ah

[potent sun]

i mean

[lucid yoke]

like research?

[potent sun]

I want to sort tags for prompting

because each model has its own method and prompting order matters in image generation

but im tired of doing it manually and I was looking for tools

[lucid yoke]

well IMO prompt order does matter slightly but not as much as you think it does

so let me get this straight

you want to feed model specific papers to a LLM so it gives you the optimal prompt order layout to you back?

like dawg, at that point its easier to compile that data into a spreadsheet and just have row 1 models, row 2 tag order and methods etc

[snow eagle]

I tried to contact support but it keeps closing my ticket 😦

[snow hatch]

Hi guys! I want to train LoRA models in local with Kohya GUI (Standalone).
I noticed that the training-help channel is hidden now.
Can anyone tell me what is the current best repo to use for Kohya GUI?
Should I use btw66 fork? Or is SD.Next better?
I'm on RTX 3060 Ti.
Also, how can I unlock training-help?
Thanks a lot! 🙏

[hollow ibex]

Hey, good evening! I'm currently using SD 1.5. I'm trying produce an image where an individual is wearing a piece of jewelry around their neck that is very similar to attached images. I however can't for the life of me seem to figure it out, I'm usually given an image of a large,thick metal collar which isn't what im looking for. I've tried: bronze/ gold ring around neck with gap/break/ opening infront, bronze/gold slender U-shaped piece around the back of the neck, necklace with gap/ opening in the front and many more. Typing in the name of the object (Torc) yields no results either. I'd appreciate advice or someone to nudge me in a direction where I can research a solution. Thank you all for your time.

[craggy marsh]

Hi everybody. Has someone installed Forge, Automatic1111 on Ubuntu ? I've succeded once but since I've reinstalled Ubuntu 22.04.5 LTS Stable Diffusion can't create the Python environment... Can some Linux expert help me ?

[cedar pulsar]

How long does it take to just the license and download information for the local model. I filled out the form on sunday and still nothing.

[lime pivot]

@hollow ibex create a lora for it, then prompt it maybe

[lucid yoke]

If you joined a "support" discord its a scam, this is the help channel

[shell ruin]

I try stable-diffusion-xl-1024-v1-0, how to deal with the edge

[lament cypress]

Hey guys im running a1111 stable diffusion but whenever u try to use flux it keeps giving me duplicated/bad results

Any ideas why?

[boreal seal]

So the prompts are odd, first this doesnt look like it’s working off the dreamshaper checkpoint? Neither does the style resemble the realistic lora model I installed? And overall like for the prompt as simple as it is, you’d sort of expect one cat on a moon right? In each generation it’s multiple.

[ornate elk]

Hey, Auto1111 doesn't support flux.
You would need to use Forge Webui or Comfyui

[vocal burrow]

base resolution is too high.

[lament cypress]

Aah, im running it on serverless runpod, calling it via api

[vocal burrow]

sd1.5 is trained for 512x512,
sd2 768x768,
sdxl and above 1024x1024

For each you can deviate about 25/30% off. But if you go too far you will get duplicate artifacts.

(you can upscale / hires.fix to whatever resolution afterwards)

[boreal seal]

Ohhhh

Interesting

Thanks I’ll try and bring it down tomorrow

[shell ruin]

I try https://platform.stability.ai/docs/api-reference#tag/SDXL-1.0-and-SD1.6/operation/masking ，Why is there black on the edge? How to solve it?

[vocal burrow]

beware of scammers contacting you in DM

👀

stability staff have the according discord roles to show for it.

[shell ruin]

Hello ,anyone help !

[vocal burrow]

not enough information, we don t know what you re trying to achieve, what software you re using, which settings, what s the input image, etc

[shell ruin]

@vocal burrow

[upper linden]

here is the entire log, its still giving me the same error: RuntimeError: mixed dtype (CPU): expect parameter to have scalar type of Float

@echo off
git pull
set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --use-zluda --update-check --skip-ort --ckpt-dir "F:\Zulda Backup\Models\models\Stable-diffusion" --lora-dir "F:\Zulda Backup\Models\models\Lora"

call webui.bat

[unkempt meadow]

Can someone help me with this ?

[upper linden]

also 7800 xt is my gpu on thje 25.4.1 driver

[tame sun]

Are AMD 9070 GFX files optimized yet?

Get this scam shit out of here.

[ornate elk]

can you screenshot your installed extensions?

the extensions folder or in the extensions tab

also make sure to whitelist the webui in any browser adblocker

[half wave]

hey is here a lora trainier for realistic ?:D

[tawny saddle]

yk

xd, it would be funny if ppl came here for general tech support

thatd be pretty annoying to staff i imagine

like, anyone know why my tv aint working, just static

xdd

[austere stag]

question im doing a board for pony can you tell me what these jumpers??? are called and how to add them cant seem to find them

[ruby yoke]

those are just expansion connectors. you don't really need them. you can stretch the noodles out as far as you need

[mystic dirge]

Forge keeps trying to free up 50GB of VRAM.... idk why, but it looks like im gonna have to rob NASA for their Computer parts now ;\

[open fox]

if you drop the line at a blank space you will be asked what to do. Simple select "reroute" and it will be created.

[pure glade]

Hey, when using adetailer I keep getting images like this, using face_yolo8n, someone has any clue why is this happening? Model used: NoobAI

[fallow knoll]

something is off with your adetailer settings i think

[vocal burrow]

most likely you copy pasted your global positive prompt into adetailer's prompt

adetailer's prompt should only describe the face.

[pure glade]

I left my adetailer prompt empty, didn't change anything, just enabled and thats it

[vocal burrow]

so yup it's doing exactly as it says.... it's doing a second pass on your face and applying your whole prompt

[upper linden]

i mean it worked before and i didnt do that at all

[lime pivot]

Hello guys, can you tell me what is wrong with my pose for OpenPose Editor? I’m trying to create a "superhero landing" pose, but I can't get the output to match

[last skiff]

Ey, anyone have seen the TG bots which lift clothes? How do i do it locally??? Any brother who can help me? I'm getting crazy

[ornate elk]

can you screenshot your txt2img settings too?

when you get the error?

[lime pivot]

@ornate elk hello bro do you know why my open pose isnt matching with the output?

[shadow heron]

so I'm using a 5800x cpu for generating ai images, would the speed of creating the images be improved if I use my gpu 7800xt instead?

lol was that spam that I just saw

[ornate elk]

hey, try set the openpose preporcessor to dw_openpose-full.
Also you can try change the resolution to a square if the figure of the pose is also more square like.

[ornate elk]

yep

[ornate elk]

yes it would be increased significantly

[lime pivot]

@ornate elk thanks ! will try it in a bit

[shadow heron]

yeah I'm using my cpu right now. cause torch wasn't detecting my gpu. I made it so that torch can cskip the gpu test. I will revert it back so I can show you why it is not detecting the gpu

or how it is not detecting the gpu bmy bad

[ornate elk]

probably because you didnt installed the webui version made for AMD GPUs

[shadow heron]

I followed the instructions here https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs

[ornate elk]

ah okay they are old and not recommended

[shadow heron]

wait the instructions are old?

[ornate elk]

yep

outdated

[shadow heron]

crap what's the most updated instructions

[ornate elk]

I made a lot of AMD Guides that are working and updated:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Webui-Installation-Guides
You want to follow the Automatic1111 with ZLUDA Guide.
That gives you the best performance/usability
If you want to use Flux models you would need Forge with Zluda.

[shadow heron]

I'll take a look

[ornate elk]

Sure, let me know if you have any questions

with a 7800xt you can generate images in seconds

im on a 7900xtx ^^

[shadow heron]

@ornate elk do I have to reinstall from scratch?

[ornate elk]

yes

[shadow heron]

I'll probl uninstall everything then. Can I just delet the webgui folder I created

[ornate elk]

yep

also check that your AMD Driver is on 25.3.1 or 25.4.1 and not newer

[shadow heron]

so I follow the "Getting the Automatic1111 AMD Fork" and then "setup with zluda" with the instructions below the important below for gpus newer than than rx 6800?

[ornate elk]

yes exactly 👍

[main gate]

any chance someone can help me understand impainting it seems like no matter what settings i use to try and inpaint and fix the hands all it does is blur them or change the color of the hands

[ornate elk]

do you have an example?

[main gate]

Yes give me just one second to get back to my pc

[shadow heron]

what error should I be seeing when running the step 7. is it the torch not seeing gpu?

[ornate elk]

yea when zluda ist not setup

[shadow heron]

seems to be stuck here

do I have to wait for a lil while?

[ornate elk]

let it run, looks okay

@woeful hemlock saw you scam noob, ban is on the way

[ornate elk]

oh it should opened up in your browser by itself

[shadow heron]

oh its opened lol didn't see it