🤝｜tech-support

[ornate elk]

depends on your hardware and webui

[cosmic furnace]

You did this ? 😮

[dense obsidian]

I'm willing to have it take as long as it needs

[old parrot]

What hardware do you have? Curious what the pros are using

[ornate elk]

i used a 60fps source video, and extracted it to 15fps,
Later i interpolated it two times and put it back to 60fps

so you have less frames to convert

[dense obsidian]

ahhh, I haven't gotten to the interpolation step yet, but that sounds like a great idea to save on rendering time

[ornate elk]

the "pros" should use an nvidia gpu (with lots of vram) haha, but in my new PC is an AMD 7900XTX ^^ and in my old PC a GTX1080

[worldly rain]

guys I keep getting completely pixelated-glitched pictures as a result, like the whole image is mixed pixels, I tried using different sampling methods, updated the extensions and restarded sd. I'm on sd 1.5, tried using different resolutions, and tried with/without loras, how do I fix this?

[old parrot]

Sounds epic! Do you use Flowframes for interpolation?

[worldly rain]

for example

[ornate elk]

i used it yes, but it was buggy a bit (because the free version didnt had an update since a year)
So im using Rife-ncc-vulkan, and made my own Batch scripts with the help of chatgpt ^^

[dense obsidian]

could I get a list of what you're using to make that video?

[old parrot]

I'll try it again though badass work!

[cosmic furnace]

@ornate elk how good your gpu handling video generation?)

[ornate elk]

i also made stuff like this xD

[old parrot]

Goals!

[ornate elk]

thx! but you should delete it in a few mins again ^^

[ornate elk]

i didnt tried AnimatedDiff, but it requires a lot of vram

also didnt used Deforum

[karmic crown]

oh well, now i saw everything (at least related to SD)
whatever i do a1111 works faster than comfy 🙈

[cosmic furnace]

But u have 24gb ? or amd vram counts less than nvidias? xD

[karmic crown]

m3 pro does not have "performance" mode like ultra 😦

[cosmic furnace]

btw dyu see the thing that some tests show as if amd gpu's use about 1gb more vram than nvidia counterparts?) in some games)

its about optimization or whuht? cant believe. I was pretty sure amd & nvidia use same vram modules... Was looking for some card reviews

or depends on whos the brand of gpu ?

[karmic crown]

it is not about vram, sotware is more optimized for nvidia

[cosmic furnace]

Y pretty sure its something to do with it 😦

[dense obsidian]

@ornate elk do you mind if I add you as a friend to dm you about this video making process? I'd love to get rolling on this

[cosmic furnace]

amd is just a lil bit behind with soft 😒 . Or overall optimization 🙂

[ornate elk]

Sure, i would recommend getting Waifu2x-Extension-Gui
it has every video feature included (also rife-ncc-vulkan for interpolating)
here it is: https://github.com/AaronFeng753/Waifu2x-Extension-GUI
i fount it after i made my scripts, so im using only rife-ncc-vulkan for interpolating:
from here: https://github.com/nihui/rife-ncnn-vulkan
For Frame extraction i use FFmpeg and made some scripts, but ill guess you can use your vlc trick or it could be included in Waifu4x Gui
Then i used controlnet temporalnet and other controlnets

[ornate elk]

sure!

[worldly rain]

can somebody help?

[ornate elk]

yes 24gb vram on AMD is not the same as on nvidia because for now we only had the directml webui on windows which was very vram unoptimized.
with Zluda its a different world.
and for sure on Linux with rocm its also another story (much better)

[cosmic furnace]

@ornate elk what is this flowframes?)

[dense obsidian]

I'll get that extension, I'll be back though, have to walk my dog for a little

[ornate elk]

Flowframes is also a Tool to Interpolate Frames.
Means a video contains for example 30 images (frames)
if you interpolate them, you get frames between every frames, so in the end you get 60 frames and a much smoother animation

[ornate elk]

yes i noticed that too

[ornate elk]

can you show your txt2img settings?

like steps, hires fix, etc, resolution

[worldly rain]

here

[maiden haven]

I have 2 LORAs, one is trained on the full body character and the 2nd is their face. I'm combining them when I generate with 1.0 weights and calling their keywords/tags in my prompt. Is this the best way to create a static character I can use anywhere, or should I be inmasking that face LORA onto the images?

[worldly rain]

i tried many different sampling methods btw

[unique bear]

Anyone know why animatediff always generates really fuzzy pictures?

[ornate elk]

can you try euler a ? as sampler

oh wait

whats the model?

sdxl?

or 1.5

maybe you load a wrong vae

[worldly rain]

sorry for the trouble, I fixed the problem. turns out I've been using an old "SDXL" VAE for an anime drawing. I'm on sd 1.5 lol

[worldly rain]

this right here lol

[cosmic furnace]

so i can do some vids on mah 7900 xt too? what zi limitations usually?)

[ornate elk]

you can do everything on it because videos are just frames, so you always can convert them back and forth

[cosmic furnace]

Why often popular yt vids is somewhat 3-4 sec lenght ? Just changing angle

[ornate elk]

because long videos contain hundres/thousands of frames

[cosmic furnace]

https://www.youtube.com/watch?v=sgM6Jj73cr8&ab_channel=JustMovies

havr you seen this?) guess y

[ornate elk]

12 seconds 60 fps vids are roughly 750 images

[cosmic furnace]

i was just wondering how far can i go...)

https://youtu.be/3oOePyuDw_o and u need more powa for smthing like this y ?

My friend said unlimited powa xDD ... Like 4090 or more ...

But hes not an expert 🙂

Just a thought)

[chilly knot]

who know 3.68s/it its normal speed generate for 1050 ti?

[ornate elk]

nope, these ae multiple generated short sequences

thats doable, still requires lot of work and editing

[haughty hare]

could I use rocm in a virtual machine running linux so I don't need to dual boot?

[cosmic furnace]

@ornate elk So yud recommend sd.next over usual directml auto. Or u cant because ur biased? 🙂

[ornate elk]

nope

i prefer directml and zluda directml over sd.next (but sd.next has the better implementation)

[ornate elk]

nope

[cosmic furnace]

why everyone try to run virtual simulation on win when its obvious fail -_-

u need full powa for sd)

[undone locust]

hi guys, noob question, i have generate a batch image and i like all iwould like to upscall all, how i can do that pleas ?

[ornate elk]

Hey, cool images! You can run then trough SD upscale script in img2img, it has a batch folder option

[undone locust]

thank :, it's crystalClearXL_ccxl model
i see we can save generate image inside a folder 🙂

[ornate elk]

they are automaticly saved in the outputs folder

[undone locust]

it's hard to get the result i want but it's in progress ^^ i try to do that. I have problem with background and size of the caractere inside scene

[ornate elk]

ouh wood cut art, nice!

[undone locust]

Not me ^^ but i try to make something in this style

[ornate elk]

ah okay

[undone locust]

my closest result

Do you have advice to acheive my goal ? for the background and always same color (wood)

[ornate elk]

hmm would recommend the add-detail lora

can give some more textures

[undone locust]

okok thank you , i'm using Forge UI , Lora is another model or is something else ?

[karmic crown]

can they use loras in forge?

i know it is limted comapred to a1111

wait, that is focus, not forge

sorry

i mixed up those two

[ornate elk]

loras are small files, that got trained on a specific style, character or other stuff. that goes on top of a model to help it generate and "know" what to generate

[undone locust]

ok i gona learn on this way ty

[undone locust]

i progress ^^ ty for help me

[ornate elk]

wow big improvement

[cosmic furnace]

nice, had a power shut down during first launch compile generating =\

[karmic crown]

this is crazy, default a1111 with my old params works the fastest on my new macbook
faster than webui-forge and faster than comfy 🙈
i tried both stable and develop versions of torch, but a1111 stills runs the fastest
enough for today, i am going to bed

[ornate elk]

Noooo

[onyx solar]

Seeing those carvings with the turquoise in them made me want to do some.

[green ferry]

hello. I am trying to download automatic 1111 on mt windows 11 computer and I install python and git and set the path, but when I click on webui-user.bat I get this message:

if anyone knows how to resolve this i would greatly appreciate it

[noble belfry]

dont install stable diffusion stuff under user, it causes all kinds of environment path issues. you're honestly better just making a C:\AI directory and putting in anything from github into that

[green ferry]

okay so would you recommend re git cloning?

or just moving it to a folder for c:ai

[noble belfry]

keep any models you've downloaded and just clean install

even if you werent getting issues now, you'd run into them later. 🙂

[green ferry]

same error. redownloaded the git zip for automatic111

[worldly comet]

no pip module found how do I fix

[distant plover]

Trying to run this Instant ID controlnet. Get an error:

[covert jolt]

健康

school

[distant plover]

school?

[upper linden]

[dapper frigate]

Guys , I am trying to upscale image using SD image2image api , but since I have file saved in the url I have to first convert it to blob and then buffer , but it returns error that content type can not be application/octet-stream but in reality its image/png when I check the buffer header. anybody has soultion to this?

[ornate elk]

Don't download it as zip

Please follow the install guide. Its in the Pinned Messages of this channel

[ornate elk]

Would need more info, python version, screenshot of the error

[undone locust]

Very nice, i'm trying carving too, how do you setup the background like that ?

[undone locust]

Question about this settings : Add extended info (seed, prompt) to filename when saving grid, where the file are stored ?
I have cheked in txt2img and txt2img grid but no file , this file save the complet prompt + settings ?

[spring glacier]

how do you add the sd vae and clip skip

[ornate elk]

To activate the VAE and clip skip menu go into
Settings, User Interface, Quicksettings, there add sd_vae, stop_clip_at_last_layer,
then hit apply and reload ui

[spring glacier]

thanks

[ornate elk]

it says to "filename" so the image will contain more information in the name

not as extra file

its limited to windows filename lenght, so not a good option, better use [datetime] as save format

the information gets stored in the MetaData of an image in any way

[undone locust]

okok , so its not possible to backup the text (prompt + settings + model + seed etc) under the generated image ?
I'm on windows is check the property of the photo i cant find the "metadata"

[ornate elk]

what do you mean by "under the image"

everything is stored in the metadata which you can only see with an exif capable tool
like Exiftool or in auto1111 the png info tab

what you can do is activate "save a json or txt file for every image which contains the infos" in the settings

[undone locust]

this one all data for generated image

[ornate elk]

yes thats what get stored into the meta data

[undone locust]

nice thank !

[ornate elk]

you can drop the image into the png info tab and then click on Send to txt2img, then it will set everything like it was for that image, sometimes you need to switch the checkpoint still

[undone locust]

yes i tried its working perfectly 🙂

[umbral plover]

What is wrong?

[quick forge]

just to clarify, if i want to download this https://huggingface.co/SG161222/Realistic_Vision_V6.0_B1_noVAE/tree/main I don't need the directories or anything, just the safetensor files?

[ornate elk]

yes

[robust vale]

anyone know how to shift all extensions from a1111 to forge ui

[trail arch]

I want to perform virtual try-on of clothes... Any suggesstions if this can be useful in any way?

[maiden haven]

What affects a safetensor/lora/models disk size, is it just number of pictures? I've been making a bunch trying to figure htings out and I'll have one that's 5gb that took 1hr to make, then last night I left one overnight, 8 hours training and it's ... 10k KB? ? then I'll have a bunch that are pretty much the same size so i assume that's image size related

i havent tested anything yet but no errors anywhere

oh cool its all black images

[waxen ore]

for comfyui, what happened to right click on group "bypass" and the selection that turns them back on?

ok, figured it out. if you lock the group so you dont accidently move it it also removes the ability to bypass the group of nodes.

[green ferry]

redownloaded python, git and git cloned with cmd and same thing. Im using windows 11 with an nvidia gpu.

[maiden haven]

that isnt a python error at all

right click the .bat file and see if windows blocked it from execution

[green ferry]

how do i see that?

[maiden haven]

right click properties

[ornate elk]

can you show the folder with path where the bat file is located?

[sick ginkgo]

in comfyui, how do i access the filename in load image batch from dir (inspire pack)?

[ornate elk]

you copy the file path to the image from the explorer

like C:\ai\comfyui... or C:\images...

[sick ginkgo]

i meant when i save the file, how to access the input filename?

[ornate elk]

hm no clue, you can ask this in #🧣｜comfy-ui too

[ornate elk]

looks good, and when you double click the webui-uder.bat ? you get that strange error?

[green ferry]

yes

[maiden haven]

Is there a spot in kohya_ss webui to put a sampler?

look at security or advanced in that file propeties pop up i forget where execution block is, its a checkbox if its on

if its not on you wont see it

[ornate elk]

yes check the security details of the folder

[cosmic furnace]

dang it a lil bit empty today in sd.next help discord 😦

so far tested generation in sd next with 1 model... but my oh my did after some time i started to get alot of black sreen -insta restart (as if i pressed reset button) espec when trying to make that img2img higher res generation -__-

[green ferry]

i see this and then i added full control but it still pops up with the same error

[ornate elk]

do you have any third party antivirus active?

mcaffe, avg, avast, etc

[green ferry]

i do not believe so

[maiden haven]

windows 11 is weird about stuff in the C:\ as well when i mod skyrim i have to be in C:\folder1\subfolder\mods.exe

[green ferry]

i turned off windows defender and double clicked the webui again and still same error

[ornate elk]

make sure Users also have full control

[green ferry]

yeah did that and same error

[maiden haven]

make a folder on C, make another folder inside of that folder, move your unzipped webui stuff in it and try it there

[ornate elk]

he has an AI folder

[polar cliff]

Any tips on the ending part of the generation process taking forever?
From what I can tell, its the VAE step that is the problem, it'll get to about 90% very quickly and then hang there for a while, up to a few minutes or even freeze.
Using A1111 SDXL, good computer. Any ideas?

[maiden haven]

read whati said

#🤝｜tech-support message

juts make a subfolder in AI and try it

[ornate elk]

maybe worth a try

[maiden haven]

might not work but its something, idk why they started doing that

[quick forge]

I may have asked this the other day, but i do believe I was blind drunk ....
Is it possible to connect ubuntu to a windows network so I can share files between them?
or possibly some way to remote control it doing a headless setup?

[ornate elk]

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

[maiden haven]

yes you need to install an SMB server on ubuntu and mount it as a share in windows, probably 3-5 commands

[green ferry]

same message

[polar cliff]

2060 super, Windows, args: --xformers --no-half-vae --medvram-sdxl

[maiden haven]

idk i'd make a new folder under your user docs and settings and try it there

[quick forge]

samba file server?

[maiden haven]

generiaclly speaking yes

any cifs/smb server

[green ferry]

idk why its saying it cant find it when the file is right there

[quick forge]

looks like there are some instructions here
https://phoenixnap.com/kb/ubuntu-samba

[maiden haven]

if youre using that server as a NAS/docker host it may make sense to just put unraid or freenas on it instead of dealing with smb config files

[ornate elk]

@green ferry can you check that?:

[quick forge]

most of the time I only understand about half of what you say lmao.
I just mean because im new to linux and its a whole new world

ill follow that tutorial and see how it goes

[ornate elk]

try the sdxl-fp16 vae fix, as vae file

[maiden haven]

there are linux control panels that give you a web GUI for doing stuff like smb shares and all that admin stuff without having to login if you want something like that

[quick forge]

im trying to sort out my models and stuff currently. previews, descriptions etc 🙂

[quick forge]

that might actually be good

[polar cliff]

This? https://huggingface.co/madebyollin/sdxl-vae-fp16-fix

[quick forge]

actually preferable tbh

[maiden haven]

@visual girder https://cockpit-project.org/

ubuntutu has one as well iirc.. forget the name

[polar cliff]

Noticeably faster now, thanks for your help

[green ferry]

this is what my comspec says

[ornate elk]

there is the problem

[quick forge]

would you suggest any specific one?

[ornate elk]

@green ferry it should look like this:

[green ferry]

do i manually write it out or browse?

[ornate elk]

you edit that variable

and remove the stuff before C:

[maiden haven]

What does LORA verification do? i cant seem to get any to pass it whatever it is

[green ferry]

YAY IT WORKS

[maiden haven]

how did your comspec var change

[green ferry]

why would my enviroment variables be set up like that? this is a new computer

[green ferry]

https://tenor.com/view/i-have-absolutely-no-idea-liz-lemon-tina-fey-30rock-i-have-no-idea-about-it-gif-21976533

[ornate elk]

something screwed it up

[maiden haven]

i dont even know what at a comspec is

[ornate elk]

a programm, driver etc

ANT_HOME xD

so its a bug 😛

[green ferry]

https://tenor.com/view/olsen-twins-thank-you-you-da-best-you-are-the-best-blow-kiss-gif-2875584550329203042

@ornate elk

[cosmic furnace]

@ornate elk did u have any problems like that(as if pc do reset button) and is there a way to find a reason? ive had to downclock gpu for directml long time ago when having same problem. After googling a bit found that smone said it goes overmuch too high with basic mhz some times

[maiden haven]

i really need to find a good SD app and stick with it i have like 10 different installs hoggging up space

im so sick of making symlinks in powershell i hate windows terminals

[ornate elk]

for symlinks use the Shelllink Extension programm

[maiden haven]

oh god yes is that a right click context button sick

i always forget you can't just create shortcut and move it lol

[ornate elk]

so your PC restarts when using SD ?

[cosmic furnace]

y

when about 20-25% progress ish

when using next.sd + zluda

and high resolution 1k + in img2img tab

[maiden haven]

does it crash when you game? gpu overheating/dying?

[ornate elk]

check the vram usage when doing that, but ill gues its just a driver crash that results in a pc restart

[cosmic furnace]

the weird thing is first it did manage to make 1536-1280 once... maybe another img style but can that be a reason? o.0

mb shud make a clean instal of next.sd? But ive kinda used reinstall command -__-

[ornate elk]

i got a full pc restart yesterday too when trying to load sdxl with zluda

but sd.next doesnt even work for me

have you tried the directml fork with zluda?

[opaque vine]

A year or two ago I paid for some credits to use Stable Diffusion on some of their web app. I can't find this web app anymore - any idea where to look? I'd like to use my remaining credits for something small

[cosmic furnace]

Welp when trying to do alot of stuff & on high image resolution he seems to doesnt handle it... and i fcking lowered power usage & mhz of videcard kinda very much to be sure its not that

but it didnt helped

[ornate elk]

Stability.ai has only one web app:
https://beta.dreamstudio.ai/generate

[cosmic furnace]

btw is it my imagination or when card overclocked it do generation faster? o.0

[opaque vine]

That's the one! Thanks!

[quick forge]

what does ema only mean?

[ornate elk]

yes it can be faster when overclocking, but also unstable

[cosmic furnace]

Y ive got an consistant 1-2 second less generation with 1024 image hmgm

well suxx then i usually have to opposite to overclocking when generating 😦

because without it for it was pc reset >_> espec when sdxl or high res. On directml)

[ornate elk]

not the best explanation but an info

[quick forge]

ok....cool, cool, cool......

I still don't understand lol

inference and fine tuning

I got that part lol

not sure what those are in relation to SD

[ornate elk]

inference = image generation
finetuning = training, making models, loras

[sick ginkgo]

feeling really dumb for asking this simple question
how do i get the filename text into filenameprefix?

NVM, figured it out.
just had to convert my widget into an input and connect the two

[cosmic furnace]

sadge 😦

[quick forge]

so can anyone tell me what specifically this is? I downloaded it and not sure what to do with it.
vae-ft-mse-840000-ema-pruned_fp16

.pt

it goes in the embeddings folder?

[ornate elk]

its a ve file, it goes into the models/vae folder
its for getting better colors, (and very minor details)

[quick forge]

i also have a safetensors file of the same thing, but without fp16

[ornate elk]

did you got the crashes only on sdxl?

[cosmic furnace]

the weird thing u know whuht? It did generated 1536-1280 img in img2img once

but just without any style and hideously looking 😄

and after that 1 successful try . Not a single try it managed to do without reboot =\

but it did once...

i didnt used sdxl yet 😦

its that fricking 1 model and VAE that comes with it 😒

🫠

[maiden haven]

Ok I'm training a model and sampl[ing every 30 steps, around step 90-120 (so 4-5 seeds only so far) the people in the LORA samples went from human to mashed up toclors to now black.. is that overtraining?

[ornate elk]

whats your it/s in sd.next?

when using 512x512, euler a, 30 steps?

prompt: cute cat, masterpiece,
neg prompt: blurry, deformed,

with an 1.5 model

[neon ruin]

I'm just getting an nvidia card at this point dude

lol

[pale hatch]

I just updated via git pull and now I get these errors. How do I fix them?

venv "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\venv\Scripts\Python.exe"
fatal: No names found, cannot describe anything.
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug  1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: 1.7.0
Commit hash: 601f7e3704707d09ca88241e663a763a2493b11a
Traceback (most recent call last):
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\launch.py", line 48, in <module>
    main()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\launch.py", line 39, in main
    prepare_environment()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\launch_utils.py", line 560, in prepare_environment
    raise RuntimeError(
RuntimeError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check
Press any key to continue . . .

[ornate elk]

sdxl works on sd.next with zluda it seems, but i understand you ^^

[ornate elk]

whats in your webui-user.bat?

[neon ruin]

even with zluda I'm just getting random crashes

seems like if I up the sampling steps, attempt to do hi-res fix, or even using more realistic/photo realistic models PC just shuts off

[pale hatch]

#!/bin/bash
#########################################################
# Uncomment and change the variables below to your need:#
#########################################################

# Install directory without trailing slash
#install_dir="/home/$(whoami)"

# Name of the subdirectory
#clone_dir="stable-diffusion-webui"

# Commandline arguments for webui.py, for example: export COMMANDLINE_ARGS="--medvram --opt-split-attention"
#export COMMANDLINE_ARGS=""

# python3 executable
#python_cmd="python3"

# git executable
#export GIT="git"

# python3 venv without trailing slash (defaults to ${install_dir}/${clone_dir}/venv)
#venv_dir="venv"

# script to launch to start the app
#export LAUNCH_SCRIPT="launch.py"

# install command for torch
#export TORCH_COMMAND="pip install torch==1.12.1+cu113 --extra-index-url https://download.pytorch.org/whl/cu113"

# Requirements file to use for stable-diffusion-webui
#export REQS_FILE="requirements_versions.txt"

# Fixed git repos
#export K_DIFFUSION_PACKAGE=""
#export GFPGAN_PACKAGE=""

# Fixed git commits
#export STABLE_DIFFUSION_COMMIT_HASH=""
#export CODEFORMER_COMMIT_HASH=""
#export BLIP_COMMIT_HASH=""

# Uncomment to enable accelerated launch
#export ACCELERATE="True"

# Uncomment to disable TCMalloc
#export NO_TCMALLOC="True"

###########################################

[neon ruin]

with stylized models I can create 50 images in like 2 minutes no problem, 7900XT is super fast

Just getting frustrated at the limitations here

[ornate elk]

thats the webui.bat, i mean the webui-user.bat

[pale hatch]

Oh .bat, I checked webui-user.sh

[ornate elk]

okay, i dont get these issues at all
seems like a hard/software issue with windows or memory management, power supply etc?

[pale hatch]

I don't have that as a .bat file

[ornate elk]

its the webui-start.bat probably

[pale hatch]

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--medvram --no-half-vae --opt-sub-quad-attention --opt-split-attention-v1

call webui.bat

[ornate elk]

okay, add:
--use-directml --skip-ort --upcast-sampling

then relaunch

[pale hatch]

Ok lemme rerun

no module 'xformers'. Processing without...
Installing requirements for easyphoto-webui
Installing requirements for modelscope
Installing requirements for easyphoto-webui
Installing requirements for diffusers
Launching Web UI with arguments: --medvram --no-half-vae --opt-sub-quad-attention --opt-split-attention-v1 --use-directml --skip-ort --upcast-sampling
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
Traceback (most recent call last):
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\launch.py", line 48, in <module>
    main()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\launch.py", line 44, in main
    start()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\launch_utils.py", line 677, in start
    import webui
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\webui.py", line 13, in <module>
    initialize.imports()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\initialize.py", line 34, in imports
    shared_init.initialize()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\shared_init.py", line 28, in initialize
    directml_do_hijack()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\dml\__init__.py", line 69, in directml_do_hijack
    import modules.dml.hijack # pylint: disable=unused-import
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\dml\hijack\__init__.py", line 5, in <module>
    import modules.dml.hijack.diffusers
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\dml\hijack\diffusers.py", line 172, in <module>
    ) -> Union[diffusers.schedulers.scheduling_lcm.LCMSchedulerOutput, Tuple]:
AttributeError: module 'diffusers.schedulers' has no attribute 'scheduling_lcm'. Did you mean: 'scheduling_ddim'?
Press any key to continue . . .

[ornate elk]

delete the venv folder and relaunch

[pale hatch]

Jesus Christ, that folder's 3 gigabytes

[maiden haven]

watching this lora model sample every few minutes is fascinating

[cosmic furnace]

it was smthing like this?)

40 steps karras 2d ++

[maiden haven]

the venv folder is your entire python environment for the app. its every single package that pip installs etc. thats why ideally oyu don't break your venv folders but since its python everone will break it anyway

i do need to set up a pip caching proxy because theres way too much duplicat downloading

[ornate elk]

thats really slow wow, did you used hires fix or loras?

[cosmic furnace]

no o.0

[ornate elk]

im getting 5 it/s on sd.next with zluda,
and 16 /its on directml (auto1111) with zluda

[cosmic furnace]

had to do downclocking as usual

mb even lots of it

was afraid that a reason

made a 1030 mv -4 or 5% powa and about 2400 mhz or 2300... but that shudnt affect that much y ?

also ive got 7900 xt not xtx)

[ornate elk]

can you try the directml fork with zluda?

i think you should get much more it/s

[cosmic furnace]

mb)

[ornate elk]

let me know if you tried it

i also found whats slowing sd.next down... for amd

[pale hatch]

venv "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\venv\Scripts\Python.exe"
fatal: No names found, cannot describe anything.
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug  1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: 1.7.0
Commit hash: 601f7e3704707d09ca88241e663a763a2493b11a
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
Installing openmim
Uninstalling mmcv mmdet... (if installed)
Installing mmcv mmdet...
Installing sd-webui-controlnet requirement: mediapipe
Installing sd-webui-controlnet requirement: svglib
Installing requirements for easyphoto-webui
Installing requirements for modelscope
Installing requirements for easyphoto-webui
Installing requirements for diffusers
Launching Web UI with arguments: --medvram --no-half-vae --opt-sub-quad-attention --opt-split-attention-v1 --use-directml --skip-ort --upcast-sampling
2024-02-25 18:11:31.057845: I tensorflow/core/util/port.cc:113] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
WARNING:tensorflow:From D:\Programme\StableDiffusion\stable-diffusion-webui-directml\venv\lib\site-packages\keras\src\losses.py:2976: The name tf.losses.sparse_softmax_cross_entropy is deprecated. Please use tf.compat.v1.losses.sparse_softmax_cross_entropy instead.

no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
Traceback (most recent call last):
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\launch.py", line 48, in <module>
    main()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\launch.py", line 44, in main
    start()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\launch_utils.py", line 677, in start
    import webui
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\webui.py", line 13, in <module>
    initialize.imports()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\initialize.py", line 34, in imports
    shared_init.initialize()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\shared_init.py", line 28, in initialize
    directml_do_hijack()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\dml\__init__.py", line 69, in directml_do_hijack
    import modules.dml.hijack # pylint: disable=unused-import
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\dml\hijack\__init__.py", line 5, in <module>
    import modules.dml.hijack.diffusers
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\dml\hijack\diffusers.py", line 172, in <module>
    ) -> Union[diffusers.schedulers.scheduling_lcm.LCMSchedulerOutput, Tuple]:
AttributeError: module 'diffusers.schedulers' has no attribute 'scheduling_lcm'. Did you mean: 'scheduling_ddim'?
Press any key to continue . . .

[ornate elk]

okay, go into the stable-diffusion-webui-directml folder
click in the explorer bar (not searchbar) and type cmd
hit enter

then type these two commands one by one
.\venv\Scripts\activate pip install diffusers --upgrade

then relaunch

[pale hatch]

Same error

fatal: No names found, cannot describe anything.
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug  1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: 1.7.0
Commit hash: 601f7e3704707d09ca88241e663a763a2493b11a
no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
Installing requirements for easyphoto-webui
Installing requirements for modelscope
Installing requirements for easyphoto-webui
Installing requirements for diffusers
Launching Web UI with arguments: --medvram --no-half-vae --opt-sub-quad-attention --opt-split-attention-v1 --use-directml --skip-ort --upcast-sampling
2024-02-25 18:20:41.943781: I tensorflow/core/util/port.cc:113] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
WARNING:tensorflow:From D:\Programme\StableDiffusion\stable-diffusion-webui-directml\venv\lib\site-packages\keras\src\losses.py:2976: The name tf.losses.sparse_softmax_cross_entropy is deprecated. Please use tf.compat.v1.losses.sparse_softmax_cross_entropy instead.

no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
Traceback (most recent call last):
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\launch.py", line 48, in <module>
    main()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\launch.py", line 44, in main
    start()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\launch_utils.py", line 677, in start
    import webui
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\webui.py", line 13, in <module>
    initialize.imports()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\initialize.py", line 34, in imports
    shared_init.initialize()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\shared_init.py", line 28, in initialize
    directml_do_hijack()
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\dml\__init__.py", line 69, in directml_do_hijack
    import modules.dml.hijack # pylint: disable=unused-import
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\dml\hijack\__init__.py", line 5, in <module>
    import modules.dml.hijack.diffusers
  File "D:\Programme\StableDiffusion\stable-diffusion-webui-directml\modules\dml\hijack\diffusers.py", line 172, in <module>
    ) -> Union[diffusers.schedulers.scheduling_lcm.LCMSchedulerOutput, Tuple]:
AttributeError: module 'diffusers.schedulers' has no attribute 'scheduling_lcm'. Did you mean: 'scheduling_ddim'?
Press any key to continue . . .```

[ornate elk]

can you run git pull again?

[pale hatch]

D:\Programme\StableDiffusion\stable-diffusion-webui-directml>git pull
Already up to date.

[ornate elk]

okay now rename the extensions folder to _extensions

[pale hatch]

Done

[ornate elk]

and then delete the venv folder again and relaunch

[pale hatch]

Okay

[ornate elk]

if that doesnt work its quicker to reinstall the webui into a new folder and move the models over

i have a easy guide in the pinned messages of this channel (not the zluda one)

[pale hatch]

Okay

[ornate elk]

whats your gpu btw?

[pale hatch]

Radeon RX 570 (8GB VRam)

I can sometimes get images.

(Somehow got a lot of funny glitches while running Cyberpunk2077 and generating an image at the same time)

Relaunching rn

[ornate elk]

yea dont game and SD at the same time xD

[maiden haven]

whats the most realistic upscale for faces

[pale hatch]

It booted up!

[ornate elk]

perfect, so it was a extension that was outdated and caused the error

now you can reinstall them, one by one

dont copy them over, better to fresh install them via the extensions tab

[cosmic furnace]

intresting...

well i guess when it all get updated gonna be better?)

weird thing is... it managed to do once a generation on 1536x1280 and once-twice 1408-1152 resolution

w.e shud try directml indeed i guess... or reinstall next sd mb

[fierce light]

Hello, I'm getting this error message while trying to use Dreamshaper XL model on automatic1111 UI

[warped ginkgo]

NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.

[ornate elk]

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

[fierce light]

My GPU is 4070Ti and my brother's is 3060. We both got the same error message. How can I look inside of ''webui-user.bat''? I double clicked it and got this

[maiden haven]

right click and edit it in notepad

[fierce light]

this

[ornate elk]

seems like you dont have python installed correctly

please take a look at my install guide. its in the pinned messages of this channel
[nvidia] #🤝｜tech-support message

[ornate elk]

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

[warped ginkgo]

Geforce RTX 4070 16gb

--autolaunch --no-half-vae --xformers

[warped ginkgo]

^

tryna gen a 1920/1080 img

[undone locust]

Hi guys, it's possible to upscale real photo 4000x4000 by my drone (not ai generated) , i try extra and img2img with TilledDiffusion or Upscaller but when i zoom +++ i dont see much difference , what is the best way to do that ?

[ornate elk]

okay looks normal

[quick forge]

which folder do bin files go?

ip_adapter.bin

[warped ginkgo]

I think it's because the Lora I was using wasn't updated to 2.1

[ornate elk]

models/controlnet

[quick forge]

ok, thought so. now I need to find where I got that one lol

found it. damn I love the search here 😄

[haughty hare]

what's the difference between these two?

[prime hinge]

Cannot locate TCMalloc (improves CPU memory usage)
Python 3.10.9 (main, Mar 1 2023, 18:23:06) [GCC 11.2.0]
Version: v1.7.0
Commit hash: cf2772fab0af5573da775e7437e6acdca424f26e
Traceback (most recent call last):
File "/home/max/Desktop/stable-diffusion-webui/launch.py", line 48, in <module>
main()
File "/home/max/Desktop/stable-diffusion-webui/launch.py", line 39, in main
prepare_environment()
File "/home/max/Desktop/stable-diffusion-webui/modules/launch_utils.py", line 384, in prepare_environment
raise RuntimeError(
RuntimeError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check

Just started happening out of the blue. Does anyone have any clue why the gpu isnt detected any longer?

[fierce light]

Is python 3.12.2 okay?

[ornate elk]

no, use 3.10.11 (just read my guide)

[prime hinge]

Best to just run 3.10 tbh

[quick forge]

Is there any recommended way of organizing model files and such

[ornate elk]

subfolders

[fierce light]

Should I delete the whole auto1111 before starting your guide? or installing python, git and adding stuff to command args is enough?

[ornate elk]

Yes delete it

Then we can be sure everything is set up correctly

[fierce light]

Can I install it on any other drive than C(system directory on mine)?

[quick forge]

I can't figure out where this pth file should go, I can't find it in any options in A1111 anywhere.
4x-UltraSharp.pth

no real instructions anywhere on where to put it either except one reference to the controlnet models folder, but I can't find it anywhere in controlnet

[ornate elk]

Yes

On any drive

[ornate elk]

upscalers go into models/esrgan

[quick forge]

oh

[neat widget]

I'm using the web api and I keep getting images that are not up to the standart I'm looking for.
Using sdxl in photography mode but keep getting quasi tiled images.
I'm usually getting 3 - 6 photos at the time and I'm not sure how this is possible since I'm getting really good quality back some other times

[quick forge]

now where do I find it in A1111 to use it? I think I've looked in every drop down menu and don't see it anywhere

oh wait i think i found it

yup 🙂

[sweet basin]

fatal: detected dubious ownership in repository at 'E:/stable/sd/webui/stable-diffusion-webui-directml'
'E:/stable/sd/webui/stable-diffusion-webui-directml' is owned by:
'S-1-5-21-430890288-2059653364-1645812397-1003'
but the current user is:
'S-1-5-21-2723868129-456139733-2995132789-1001'

how to change ownership?

I reinstalled windows

[quick forge]

stable cascade needs 24GB?!?!?!!

[ornate elk]

did you moved the webui ?

where is it located now?

[sweet basin]

same location

[ornate elk]

is the E an external drive?

[sweet basin]

ssd m.2

[ornate elk]

ok, then right click the stable folder
and under properties, security you give your user, and system all priviliges

[fierce light]

My brother got this error message while trying the last step(save and launch webui-user.bat) of the guide

python was successfully installed though

[sweet basin]

NOTHIGN

[ornate elk]

did he checked "add python to path" ?

[fierce light]

yes

[ornate elk]

then he should open up a cmd and type
where python

that should show a correct path to 3.10.11

[green ferry]

do you guys know what triton is?

[fierce light]

He did it and launched webui-user.bat again, he got same error again

maybe restarting computer could fix idk

[ornate elk]

he should run the command and show me what the output was xD

that command doesnt fix anything, its for checking if python is installed correctly

[fierce light]

sorry, ok he'll send it now

[ornate elk]

np ^^

[ornate elk]

okay seems like he also installed python via the windows store, he should uninstall that

then run the python 3.10.11 installer again, and then click modify, next, make sure "add python to system variables" is checked

[mild sandal]

how can I uninstall?

[ornate elk]

via windows settings, apps

or system control

[mild sandal]

both?

[quick forge]

I just reformatted an HDD to BTRFS and now I can't write to it, like I want to copy/paste some stuff there but no paste option comes up.

[ornate elk]

hmm are there only these two?

[mild sandal]

yes

[ornate elk]

hmm okay

then you can do something different

open up the cmd again and type
where python

then copy the second path (not the one with windows store)

then edit the webui-user.bat
at the line PYTHON=
you paste the path you copied

then save and launch

[sweet basin]

should I delete everything and reinstall new one

[ornate elk]

would be easier

you can follow my guide

from the pinned messages

[sweet basin]

okay

[ornate elk]

and safe your models and stuff

you can copy them over

[ornate elk]

@cosmic furnace
for sd.next go into settings and make it lok like this for max performance:
then hit apply and restart

[winged pelican]

Hello - I'm trying to use stable-cascade on Linux, and I'm only able to generate low-quality images with dead eyes. I also tried ComfyUI, and I get results like this with the default workflow. Only slightly better with 1024x1024 image size. Can anyone please spot what I'm doing wrong, or share a basic workflow that I can use for testing?

[quick forge]

so apparently I need to reinstall ubuntu due to some permissions I fked up..... already y webui's on a different HDD than the boot drive.... is there anything I need or could save to make reinstalling and getting SD running again easier?

[winged pelican]

~/.cache/huggingface
and don't forget to backup any bookmarks.

[quick forge]

There isn't much on this computer, it's basically only for SD

I don't seem to have a cache folder

nvm i think i was looking in the wrong place

[winged pelican]

Unlikely. If you're looking in the gui, it might be hidden. ctrl+h maybe to unhide?

[winged pelican]

This test worked! https://gist.github.com/comfyanonymous/0f09119a342d0dd825bb2d99d19b781c

[quick forge]

how is saving that folder going to make installing stable diffusion easier?

[winged pelican]

Depending on how you have it loaded, you might not have to re-download gigs of files.

[quick forge]

all of the gigs of files are on a separate HDD. all of my WebUI's and stuff are there with all the models and everything

the entire cache folder is only 7GB

[maiden haven]

its only 7gb which is fine until its your 8th time downloading it

resolving a broken venv for 90% of the people in here is deleting venv and repulling.. everything

[maiden haven]

geez adetailer makes my faces ugly

[noble belfry]

tbh adetailer is best used in images where the face isnt the most prominent thing and the actual model did a half hearted attempt at rendering it.

[maiden haven]

ah ok, never used it. the model for it does seem to make a huge differnece

this face/subject detection is rad i need to figure out what to do with it

[noble belfry]

i mean it can do great stuff, but i wouldnt consider it a core thing in any process i use. but depends

[ornate elk]

What's your GPU?

[maiden haven]

4090

[ornate elk]

Hmm

Adetailer should make a huge difference

[maiden haven]

yea its interesting, kinda good sometimes kinda horrible other times. it's making certain parts of the fcae a little too distinct, like this character looks like squidward butt chin with it on but not without it on

[amber silo]

Oops, sorry

[cosmic furnace]

Reahly? ;O

[ornate elk]

yes make sure to set these settings like me:
#🤝｜tech-support message

should be much faster

[cosmic furnace]

indeed i have lil bit weak PCU (650w) for an 7900 xt buut...in games & directml no problems yet... hmgm

and downvolting to 1025 & -10% power wats usage didnt helped avoid this reset >_> . Havent yet tried fresh instal of sd.next or directml zluda yet)

[ornate elk]

650w 😮

[cosmic furnace]

yeahh))

was thinking about getting smthing bigger soon >_>

and was wondering if it ll be enough

[ornate elk]

850 would be good

[cosmic furnace]

i was told to lower even more but u cant go lower or higher than 10% 😄

7x series got hardcoded in vbios or even hardware level

so ppl cant overclock them with power that much as 6x >_>

sadge...

well... directml was working just fine... with downvolting & sligh power & mhz lowering 🙂

ive read smwhere other ppl also had that problem on amd because directml made a huge jumps in mhz and resulted in insta reboot

cuz bad drivers-compability & all that

for sd directml

[ornate elk]

Okay, didn't encountered that so far on all the amd systems I tested (with directml, without Zluda)

[quick forge]

im doing a fresh install for.....reasons.... which version of rocm should I be installing and using?

ubuntu 22.04

[ornate elk]

Rocm 5.7 I guess

[cosmic furnace]

So u think all this insta resets is power probs? mostly? or can be anything

[maiden haven]

psus rarely work halfway

i mean if you wnt to tempt that just throw on cpuz or some gpu benchmark thing and let it run awhile. i have a killawatt and my 4090 + entire desktop doesnt consume over 450w i dont think, maybe under 400

[carmine thorn]

Having an issue with Stable Diffusion Web UI all of a sudden. Haven't generated in a month, load up the Web UI with no changes to hardware and image generation is taking 20x longer all of a sudden.
When generating with no prompt, the image is generated as normal, e.g. 17/ it/s and CUDA being used at 90-100% on task manager. However, when adding ANY word to either positive or negative prompt, image generation starts taking ~20x longer, with it/s dropping to a consistent 1.0/s and CUDA usage dropping to 20-25%.
0 errors are being thrown in terminal
(4080 GPU 16 GB VRAM)
What I've tried:
Disabling memory fallback on NVIDIA control panel
clean re-install
restarting computer multiple times

Would greatly appreciate any help. Thanks

[maiden haven]

i have absolutely no idea what tensorboardi s telling me lol

[ornate elk]

You can check the power draw in the AMD adrenaline software.
But for example my GPU 7900xtx could take up to 420w max.
So I think when yours is doing maxed out SD stuff the 650w can't provide enough power anymore.
AMD GPUs draw more power than nvidia ones.
Use this PSU calculator to check what is the recommended power supply wattage for your System:
https://www.bequiet.com/en/psucalculator

[ornate elk]

Hey, is your GPU driver updated?

Also pls enable memory fallback again

And then check what's inside your webui-user.bat

[carmine thorn]

drivers are all updated. I re-enabled memory fallback, still the same issue. Here's what I have in my webui-user.bat:
@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--xformers --no-half-vae

call webui.bat

[quick forge]

i keep running into errors when trying to install rocm and get stable diffusion running again.
ERROR: Cannot create report: [Errno 17] File exists: '/var/crash/amdgpu-dkms.0.crash'
Error! Bad return status for module build on kernel: 6.5.0-21-generic (x86_64)
Consult /var/lib/dkms/amdgpu/6.1.5-1609671.22.04/build/make.log for more information.
dpkg: error processing package amdgpu-dkms (--configure):
installed amdgpu-dkms package post-installation script subprocess returned error exit status 10
Errors were encountered while processing:
amdgpu-dkms
E: Sub-process /usr/bin/dpkg returned an error code (1)

[maiden basin]

Hey guys, i am thinking to use canny controllmap to change the structure landscape background of a room but it somehow doesnt work as i imagine to do so any idea why this is happening?

i inpainted the area above the long screen

[quick forge]

fuck it ill reinstall ubuntu again lmao

would model files be better on an SSD or would it make any difference at all on an HDD vs SSD for the model files and stuff.

[maiden haven]

ssd

[lost vale]

hi guys, I'm trying to run inpaint-anything on my mac mini M2 chip and I'm getting this error

loc("mps_not_equal"("(mpsFileLoc): /AppleInternal/Library/BuildRoots/4e1473ee-9f66-11ee-8daf-cedaeb4cabe2/Library/Caches/com.apple.xbs/Sources/MetalPerformanceShadersGraph/mpsgraph/MetalPerformanceShadersGraph/Core/Files/MPSGraphUtilities.mm":253:0)): error: 'anec.not_equal_zero' op Invalid configuration for the following reasons: Tensor dimensions N1D1C1H1W306252 are not within supported range, N[1-65536]D[1-16384]C[1-65536]H[1-16384]W[1-16384].
loc("mps_select"("(mpsFileLoc): 

anyone seen anything like this before?

Trying to use segment-anything extension using automatic1111 web-ui
followed the guides:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Installation-on-Apple-Silicon
https://github.com/Uminosachi/sd-webui-inpaint-anything

[unique bear]

Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

what does this error mean

Solved

[tranquil token]

Anyone has experience with prompt scheduling video workflows for stable video?
All my videos generate like this all the time.

[ornate elk]

Is your webui updated?
What does the cmd shows? Any errors when launching it?

[rough canyon]

I'm pretty new to stable diffusion, been experimenting with WebUI forge and some models, currently mainly using "AldeboBaseXL" but I keep getting bad hands. Is there a guide somewhere to fix hands/get better hand generation with AldeboBaseXL (I suppose this is based on SDXL). I pretty much know how all this works by now, but the hands are still hard and some LORA's I checked seem to only work with 1.5

[upbeat crane]

are you using any negative prompts by chance?

[sour flare]

how can I install stable diffusion?

[ornate elk]

Check the Pinned Messages of this channel for guides for nvidia and AMD GPUs

[zealous leaf]

Hello, I intend to use stable zero-123 for commercial purposes, the company has StabilityAI membership, however, I was wondering if there are any solutions that will let me use the model via API. I have found the code that will probably be used for inference, but we don't have the necessary hardware (20GB+ VRAM) for testing the solution for our case.

The next question is whether stable zero-123 capable of generating views with predefined camera angles, without generating all possible angles and movement like zero-1to3 here` https://huggingface.co/spaces/cvlab/zero123-live ?

I will be grateful if you answered those questions, preferably in reply to a message, so I will get a notification.

[dense obsidian]

A few questions I had If anyone would be so kind to take a crack at them.

If my inpaint sketch doesn't show up, or my image for inpaint sketch doesn't show up, is due to my browser, or is it something code related?

I noticed during stable ui boot up that it says my pip needs to be updated, could that be it. If it is, would updating pip cause any problems to my stable diffusion I have now? (It's version 1.7)

[upbeat crane]

it might be worth trying to load the webui in a different browser just for testing purposes, could possibly be an extension interfering so you might check and see if there are any non built-in extensions you could possibly disable and test

if send to img2img isn't populating you could always go find the image and attempt to drag-n-drop it

@dense obsidian

[dense obsidian]

Thank you! My opera gx ui is verified bugged, I just don't know the full extent of it.

I shall test in Firefox when I get home.

[ornate elk]

Make also sure the webui is whitelisted in any adblocker.
That caused issue with img2img in the past

[quick forge]

Is there a step by step guide to installing stable diffusion on Ubuntu Linux 22.04 with an AMD RX6800.
I've already had it running, I already have A1111 installed on a separate SSD that isn't my boot/linux drive.
I just reinstalled Linux due to a screw up and I think I just need to install all of the pre-reqs, and ROCm and stuff.
Just not sure the best/easiest/fastest way.

amd ke[t running into errors lastnight

[prisma quartz]

does your sd-forge eat a lot of ram???

[quick forge]

me?

[solemn moon]

any bro's wanna try help me fix my lcm

[valid flame]

guys please someone help me out to get this type of results
https://www.instagram.com/p/C3cuSRpvGOX/

aiart.mania

Reyansh as spider man.

Transform your sweet memories into visual spectacle.
Did you miss the photoshoot for the past memories for your kids or for you or for your family.
Dm if your interested to get for your loved ones.

#spiderman#marvel#spider
#aiartmania #aiartportrait #aiart #aiphotos #AiPhotography
#aiimages #aiartwork #aistillsCustom...

Instagram

[solemn moon]

keep getting eww results

[manic snow]

have you adjusted your sampler and cfg scale?

[solemn moon]

faceswap app store

[solemn moon]

i renamed the lora xl file

[valid flame]

No not bro

[manic snow]

ah, i see the problem. you're using the sdxl base model, rather than an sdxl lcm model

[solemn moon]

okei i under

[solemn moon]

thanks you sir

[manic snow]

you're welcome!

[valid flame]

can u solve my problem

[solemn moon]

ReActor extension

[manic snow]

or img2img inpainting

you change the style with a full img2img generation, then add the spider man costume by inpainting

[mighty topaz]

Someone tell me whats happen?

[crisp geyser]

i know it's pretty weird to say that but... is it actually wrong to run stable diffusions on 4gb of ram on the pc? i mean the whole memory pc not the gpu, i know it's not possible but if it was please tell me how can i limit the ram usage, thanks.

[ornate elk]

Hey, what's your GPU?

[stone spire]

hey guys, i need a help

[crisp geyser]

gt 730 4gb vram lol

[stone spire]

Failed to create model quickly; will retry using slow method. how can i fix it?

and this one Interrupted with signal 2 in <frame at 0x000001640161E050, file 'threading.py', line 324, code wait>
Terminate batch job (Y/N)?

[ornate elk]

SD won't work on that old GPUs. You need at least a GTX card.
But SD can work on CPU mode, which is very slow, but your ram is very limiting so it may dont even can be installed

[ornate elk]

Hey, what are you trying to do?

[stone spire]

I'm trying to install the dreamsshaper model I downloaded. but it never load

[ornate elk]

What's your GPU, what's inside your webui-user.bat and what's the Dreamshaper model size

[stone spire]

i have rtx 3060 6 gb vram

[crisp geyser]

i actually installed it and everything ,i'm trying to use the wd anime waifu diffusion (i'm not a weeb istg) but it seems like when it chrome starts it just uses much ram from my computer, sooo yeah i'm having that issue now

[mighty topaz]

noone know?

[ornate elk]

You need to make sure your using only 2gb models. Like Dreamshaper v8 for example.
Then you need to edit your webui-user.bat and there at the line COMMANDLINE_ARGS=
you need to add: --lowram

[ornate elk]

Can you answer the other questions too?

Would help me, to help you xD

[crisp geyser]

great, thanks for help alot

[ornate elk]

NP, --lowram makes so that it will load the model into the gpu vram instead of ram.

Maybe with that trick it will work

[stone spire]

When the model loading is completed, it returns to its previous state. When I reinstall it, it goes back to its previous state. It's like a cycle.

[ornate elk]

Okay so this model is a bit heavy for a 6gb vram GPU.
You need to right click the webui-user.bat and edit it.
At the line COMMANDLINE_ARGS=
You need to add: --xformers --medvram --no-half-vae
Then save and relaunch

[stone spire]

is it true, right?

[ornate elk]

Yes looks good

Also how much RAM do you have?

[stone spire]

I'm sorry I'm ignorant on these issues, I'm making you busy.

32 gb

[ornate elk]

Okay 32 is perfect

Then try to load the model again

Make sure you always launch the webui-user.bat and not the webui.bat

[stone spire]

oh

you're the best!

thx a lot

[ornate elk]

No problem 🙂

[stone spire]

Is there a setting you recommend for the SD to work stably?

[ornate elk]

It should work stable by default. Make sure other programms in the background are closed.
For example wallpaper engine and epic games launcher can cause stutters and issue when running SD

With the performance adjustments of your webui-user.bat it should also generate much faster with 1.5 models now

[stone spire]

and currently it does not generate images.

[ornate elk]

Oh, can you show the whole cmd when starting it?

I guess your webui might be outdated

[stone spire]

whole cmd? all of them?

[ornate elk]

Okay thats enough

@stone spire Your torch version is outdated

To update it you have to add
--reinstall-torch
To the webui-user.bat
When its done you need to remove it after that again.

[royal rapids]

idk if this is the correct place to ask but, how do i downscale photo's? I think i must be missing something but im not too sure

[manic snow]

what you can do is go to img2img and set the resolution manually and set the denoising strength to 0. alternatively, you can go into ms paint and resize it there

[solemn moon]

bruh

[manic snow]

what model is that?

[solemn moon]

im going backwards 😂

smaller 1

[royal rapids]

like so?

ms paint?

ooh

oooh

[stone spire]

okay, im trying

Do I need to reinstall extensions

[royal rapids]

UnboundLocalError: local variable 'h' referenced before assignment, imma just use paint

[manic snow]

try a model trained by the community. you can find plenty on civitai. i reccomend this one: https://civitai.com/models/4384?modelVersionId=252914

[mellow apex]

hey anyone who can help me understand how to do inpainting with automatic11111 in rundiffusion – Using juggernaut XL?

[ornate elk]

Nope, but check for updates in the webui under the Extensions tab, after that.
And dont forget to remove the --reinstall-torch
After its done

[stone spire]

okay, thx

[solemn moon]

will try thanks, was kinda hoping for xl, also, added these args but then the command window closed so removed them

[manic snow]

there are also SDXL lcm models

[solemn moon]

yea im looking now o0

[ornate elk]

When in inpainting, mask the subject you want to change.
Set the Inpaint Area to "only masked"
Then set the resolution to 1024x1024 (because your using an sdxl based model), also that's not the output resolution.
Then set the Denois to 0.5 to test how much it changes. Adjust that slider to your liking.
Also set steps to 30 and sampler to Euler a or DPM 2m karras

[stone spire]

i guess it's not installing

[mellow apex]

hmm, so i dont need to load a specific inpainting model for XL? cause some models in rundiffusion says "inpainting" on them

[ornate elk]

You can do that for more accuracy but normaly it should work with the same model your generated the image with

[ornate elk]

What's your python version?
You can check that by running
Python --version
In a cmd

Or look it up in the system control

[stone spire]

0.0.16

[ornate elk]

Thats xformers mostly but okay.
What you do now to fix everything.
First remove --reinstall-torch from the webui-user.bat.
Then save.
Download Python 3.10.11 64bit installer from python:
https://www.python.org/downloads/release/python-31011/.

Then delete the venv folder.
And then relaunch the webui-user.bat

It will setup everything correctly and should startup.

It can take a while it recreates the venv folder then.

[stone spire]

okay, it works great and i have one question. Where are our images stored?

[ornate elk]

Perfect, they are stored in the outputs folder

You can click the folder icon below the image to open it

[fierce light]

is interrupting while generation is bad thing?

[ornate elk]

Nope,

It will stop everything, depending on what step it is it can take a bit

[crisp geyser]

So, i installed everything and its running perfect, ram usage is low and everything is working fine, but when press generate it says "RuntimeError: "LayerNormKernellmpl" not implemented for 'Half'"

[solemn moon]

this for you sir

[manic snow]

:D

[solemn moon]

took long though

a yellow cat wearing glasses, soul cat, lora:loraxl:2.8
Steps: 3, Sampler: LCM, CFG scale: 1.5, Seed: 3493960838, Size: 1024x1024, Model hash: 4496b36d48, Version: v1.7.0-539-gc8a5322d

Time taken: 10 min. 31.5 sec.

[maiden haven]

jesus 10 minutes

[solemn moon]

vs what i used to do o0 maybe 50min +

[maiden haven]

i dont think i could be into this hobby without a 4090, the 1 minute imagee creation and hour long training is annoying af to me lol

burn an hour an get black images

[ornate elk]

You need to add --no-half to the webui-user.bat

[crisp geyser]

I should restart too right?

[solemn moon]

also think its my args finally catching up to me, or maybe 10min is normal for a 1660ti 🤔

[ornate elk]

If you have --xformers --medvram --no-half your fine

But sdxl will run slow on that cards

[solemn moon]

--xformers --precision full --no-half --lowvram --opt-split-attention --always-batch-cond-uncond --api

[ornate elk]

Remove --opt-split-attention and --always-batch-cond

And try --medvram instead of --lowvram

[solemn moon]

k i'll try

[ornate elk]

Also remove --precision full

It should work without it

[solemn moon]

--xformers --precision full --no-half --medvram --api

so thats it now

[solemn moon]

and this

[dense obsidian]

I have to stop working on my video late at night, can't remember what I learned the last night XD.

So if I want to keep the pose of the input, but change the out to my lora. A high cfg scale, but if I increase my denoise, won't that mess with my cfg scale, potentially ruining the input's pose?

[vestal torrent]

how do i get stablitity api (sdxl) to generate hundreds of images of the same species of animal, posing as portrait from the chest up in cartoon detailed style. Thhe prompts have random values injects for different traits like eye coor and facial expression. However, images that have 4 trais, eye color, facial expression, eyewear, and headwear, all generate to the same style and pose which is perfect, but all promots without headwear and eyewear that only have eye color and facial expression always generate in a much different style and pose making the images unusable for my purpose. How can I get them to always generate the same style and pose but with different hats, expressions, clothes, etc? I am using the same seed for every image.

[solemn moon]

a yellow cat wearing glasses, soul cat, lora:loraxl:2.8
Steps: 3, Sampler: LCM, CFG scale: 1.5, Seed: 2289362390, Size: 1024x1024, Model hash: 4496b36d48, Version: v1.7.0-539-gc8a5322d

Time taken: 23 min. 5.1 sec.

[manic snow]

are you using your cpu?

[solemn moon]

gpu surely

[ornate elk]

What was the model?

You dont need LCM sampler if you just use the LCM lora

[solemn moon]

dreamshaperXL

[ornate elk]

Dreamshaper should have LCM included

I tought

So it should work with any sampler

[solemn moon]

k let me try

a yellow cat wearing glasses, soul cat, lora:loraxl:2.8
Steps: 3, Sampler: Euler a, CFG scale: 1.5, Seed: 2481435276, Size: 1024x1024, Model hash: 4496b36d48, Version: v1.7.0-539-gc8a5322d

Time taken: 6 min. 35.7 sec.

[fair oxide]

Hello, im using depth anything + segmentation to place image into a dresser in kids room as a mockup. But whenever i generate, it doesnt keep original image and changes toys. Model im using is Realistic Vision V6.0 B1: https://civitai.com/models/4201/realistic-vision-v60-b1.

[rough canyon]

can anyone explain how I can install "photomaker" on webui forge?
can't seem to find info how to do this? it's "included" in forge, but I think I need to download some extra files to use it

[dense tapir]

Hey guys! Please drop me a quick little tutorial on how to have 2 folders which SD can load Checkpoints from! Thanks!

[dense tapir]

Try enabling Pixel Perfect

And then click the little explosion button next to the "Model" section which can be seen on your screenshots

[ornate elk]

Hey, which webui are you using?

[dense tapir]

A1111

Local

[ornate elk]

You can add a models path with --ckpt-dir "path/to/models/folder"

It will still search in its local models folder but also in the one added with path

So you have 2 different directorys

[dense tapir]

I already did that!

It didn't work

For some reason

[ornate elk]

You need /

Instead of \

[dense tapir]

Thanks!

[vestal torrent]

can someone tell me abotu controlnet

[gray oyster]

I am currently installing kohya_ss but I am unsure which cuDNN version to install, should the cuDNN version be compatible with the CUDA version I have installed or the CUDA version PyTorch is being installed with?

I have CUDA 12.1 installed, kohya_ss installs PyTorch with CUDA 11.8 support and now I don't know what cuDNN version to install

should I go with 12.x or 11.x?

[cunning matrix]

can I convert this https://civitai.com/models/118418/negativexl to another format so older versions of automatic1111 will know how to load it?

I'm using automatic1111-directml which I think (from commments on this and other textual embeddings) was forked from an old version that cannot load .safetensors embeddings

[karmic crown]

are you sure that you put file in the right folder?

[cunning matrix]

yes, I can even force loading it by mentioning in in the prompt but it throws an error

doesn't show in the embeddings list UI either

   File "F:\StableDiffusion\stable-diffusion-webui-directml\venv\lib\site-packages\torch\nn\functional.py", line 5334, in multi_head_attention_forward
        attn_output = scaled_dot_product_attention(q, k, v, attn_mask, dropout_p, is_causal)
    RuntimeError: Expected attn_mask dtype to be bool or to match query dtype, but got attn_mask.dtype: float and  query.dtype: struct c10::Half instead.

looks like it's getting loaded wrong

if I remove the embedding term from the prompt it runs fine

might be an fp16-fp32 mismatch, too? I suppose

[karmic crown]

it is is embeddings folder, right? and the file name is negativeXL_D.safetensors?

[cunning matrix]

yes

https://github.com/lshqqytiger/stable-diffusion-webui-directml this is the fork that I'm using

[karmic crown]

#🤝｜tech-support message <- check this guide

and see if you have something diffrent

CS1o might help more, I am Mac guy

but it should work without a problem as far as i know

[ornate elk]

did you downloaded A or D ?

[cunning matrix]

interesting, with the guide parameters it works. my previous ones were:

set COMMANDLINE_ARGS=--use-directml --medvram-sdxl --opt-sub-quad-attention --no-half-vae

vs. the guide's

set COMMANDLINE_ARGS=--use-directml --skip-ort --medvram --opt-sub-quad-attention --opt-split-attention-v1 --no-half-vae --upcast-sampling

[cunning matrix]

D

[ornate elk]

maybe --upcast-sampling has fixed it

[manic snow]

controlnet openpose randomly stopped working with the error "TypeError: 'NoneType' object is not iterable". i have no clue how to fix it

[ornate elk]

1.5 or sdxl?

[manic snow]

xl

[ornate elk]

hmmm

did you changed the input image?

updated controlnet?

[manic snow]

i changed the input image, and didn't update controlnet. other modules are working fine

now that i think about it, i don't think i've used openpose with sdxl before, so there's a good chance it didn't work to begin with. currently trying with a 1.5 model

yup, 1.5 works. guess i'll just stick to that then

[cosmic furnace]

Are GPU power usage in SD img resolution dependant? o.0

[ornate elk]

yes, i guess so, higher resolution, more vram gets aquired, more power consumption probably too

[cosmic furnace]

but directml shud had things more terrible with it than zluda-torch ?) or not?

I mean on zluda -sd. next never seen it uses as much vram even higher resolutions

while it seems directml always uses almost all possib vram =\

well if he doesnt lie about it ofc)

[ornate elk]

directml uses the max vram but hasnt the performance like zluda

so zluda is more demanding for the hardware as it can be "really used"

[sweet basin]

halp

[karmic crown]

unless they enter the twilight zone, as it happened to me with my new M3 PRO Macbook
you know what happened with "faster" environments 🙈

https://global.discourse-cdn.com/woot/original/4X/6/8/f/68f86446cfc1c369dd55786653c8f02e4041771a.gif

[sweet basin]

help

[karmic crown]

looks like you messed up your moduls folder

[sweet basin]

I just installed sd

[karmic crown]

open cmd, go to your stable-diffusion-webui folder and type

git status

and show me the result

also try to install module like this (be sure you are in the stable-diffusion-webui folder

venv\scipts\pip install jsonmerge

jsonmerge is an external library, and not a part of a1111 as i though, my mistake, so this thing with pip install should fix that

[brazen flare]

Hello ^^ I tried to install SD with automatic1111 but I got a torch error.

Here is the issue

Someone would know how to fix this ?

[noble belfry]

do you have a gpu?

[ornate elk]

hey hwats your gpu?

[brazen flare]

RTX 3070 ^^

the thing is that I already had SD on this computer and I had no problem. But when trying to reinstall it does not work anymore 😦

[noble belfry]

but thats AMD torch it was downloading

[brazen flare]

ooow

Will try to download webui again

Let's see

[noble belfry]

i wonder if it defaulted to that after those cuda118 peer resets at the top of the install

[brazen flare]

Downloaded again under NVIDIA webui but still the same

I also disabled my antivirus and firewall

[noble belfry]

if its helpful, the webui-forge fork of a1111 has a zip that works out of the box. functionally identical to a1111, except with more things, things are done better etc.

[brazen flare]

Niiice thanks

Will try right now ^^

[noble belfry]

https://github.com/lllyasviel/stable-diffusion-webui-forge

its by controlnet guy 🙂

half way down the page should be >>> Click Here to Download One-Click Package<<<

iirc its about 1.8gb

[brazen flare]

Super cool !! 👍 thanks a lot 😄

Will try

[karmic crown]

on my m3 PRO Macbook it works SLOWER than original a1111, even it supposed to work faster 🙈

[brazen flare]

Shall I put the sd_xl_base in models folder (like a1111) before update and run.bat ?

[karmic crown]

use some decent model from https://civitai.com/ instead of default models

[karmic crown]

just keep in mind that is a version for nvidia, and not amd

[brazen flare]

Okay nice ^^

Thank you a lot

You are so kind

[noble belfry]

well, if you dont have a model in your folder it will just download i think the 1.5 version of realisticvision with vae by default.

[brazen flare]

Niiice thanks ! It works well ^^

[reef smelt]

Hey I'm trying to use sdxl turbo but the output quality and speed don't match

[meager carbon]

hey I'm trying to generate images over 720p on an amd gpu, but it keeps saying out of memory... what can i do?

24gb 7900xtx

[reef smelt]

is that your vram?

Do you have all the necessary drivers?

[ornate elk]

whats in your webui-user.bat?

i have the same card and can generate 1024x1024

also Full HD with hires fix

[meager carbon]

i wanna generate 2560x1440

hires fix would be fine to get to this

[ornate elk]

with 1.5 or sdxl?

[meager carbon]

xl

[cosmic furnace]

is this fastest ?)

or whuht yud recommend?

[rough canyon]

Probably stupid question, but I tried using canny and got the following error : TypeError: 'NoneType' object is not iterable

Any idea why? I loaded a picture of me and my friends with snowboards, did a pre-process, I see the black/White outline but when trying to generate an image this error appears

[ornate elk]

have you disabled fp32 and no-half?

also activate upcast sampling

[ornate elk]

then you need to generate FullHD with hires fix

[meager carbon]

that's fine

but how

[ornate elk]

and then upscale that in img2img with sd upscale script

[meager carbon]

hmm

ill try...

[meager carbon]

what should i start with, like waht should the upscale factor be?

2x?

[ornate elk]

yes 2x

what you can do too:
install the new zluda backend for a webui, which can generate at higher resolutions, (without oom)
the directml webui doesnt work with sdxl, but sd.next can work with it

[meager carbon]

I see

[ornate elk]

i have a install guide in the Pinned Messages for it, if you want to try, it wont affect your existing webui, so you can test it

[meager carbon]

cool

ill try it then

[ornate elk]

yea, it brings my 7900xtx from 4it/s to 16 it/s

you can do upscaling in like 20 seconds

[meager carbon]

uhm

no worki

[ornate elk]

hires steps = 10 is needed

[meager carbon]

also just realized

[ornate elk]

now you need to restart before trying again (to clear the stuck vram)

[meager carbon]

i didn't select a upscaler

[ornate elk]

also here is a quick info/guide by me on how to generate with hires fix without out of memory: (for directml webui)
https://github.com/lshqqytiger/stable-diffusion-webui-directml/issues/373#issuecomment-1936982157

[meager carbon]

thank you

still no :/

big F

[ornate elk]

make sure you dont have any other programms open

like wallpaper engine

[meager carbon]

i don't wanna have to close a bunch of things...

[ornate elk]

and then checkout the link, there i mention the Tiled Diffusion extension which can help to get not "out of memory"

[meager carbon]

cool, ty

[ornate elk]

no problem, if you have the settings like in the image there, with tiled vae enabled, it should work, even when stuff is open, (not heavy stuff)

then if oyu have time take a look at the zluda guide, its blazing fast without the vram limitations (but only for 1.5)

[meager carbon]

im doing that now haha

[ornate elk]

let me know if you have any issues or questions, im on for an hour

[meager carbon]

you're the best

[meager carbon]

uhm

in your guide

Now go into the stable-diffusion-webui-directml folder it created what is "it"

[ornate elk]

sry for my bad english haha

"that got created"

[meager carbon]

by what though

[ornate elk]

will edit it

by git cloning

git clone a repo will get you a new folder

[meager carbon]

ohhhhhh

im silly

sry

[ornate elk]

Hey, np, the sentence with "by it" was a bit confusing xD

[cosmic furnace]

y set up as u recommended in jpg) . also trying this "experimental" speed up

[ornate elk]

I tested it and it gave me 28 it/s but sdxl and VAE files won't load with these settings in SD.next

[cosmic furnace]

wait no halv shud be enabled or disabled?)

[ornate elk]

Disabled

And fp32 disabled

[ornate elk]

Fp16 enabled

[meager carbon]

big f

[ornate elk]

:0

[ornate elk]

Only upcast-sampling enabled is needed

The other ones will slow down

[ornate elk]

Is the latest AMD driver installed ?

[meager carbon]

uhhhhhhhhhhh

[ornate elk]

24.1.1

[prime bluff]

is there any video guideline to setup stable diff in pc? I have tried using some articles but didnt work out

on nvdia card

[ornate elk]

Lol okay, 24.2.1 is not needed for that so dont update

You can update in a few days

[meager carbon]

kk

[ornate elk]

@meager carbon open up a cmd.
And run
SFC /scannow
After thats done and fixed windows errors, you need to restart the PC and then try installing HIP again

[meager carbon]

no.

[ornate elk]

Hey, checkout my guide in the Pinned Messages of this channel.
Its pretty straight forward. Make sure you use the right one for your GPU (either nvidia or amd)

[meager carbon]

my windows is really modified haha, and reversing all of that would be aweful

sadly

ill just restart my pc

see if that helps

[prime bluff]

thanks but i can only see text based guide, do you have any video recording?

[cosmic furnace]

Its weird but it looks like next sd doesnt download 1.5 model o.0

why? o.0

[karmic crown]

all pinned guides are written to be as simple as possible and easy to follow
mac users can use my script instead of my guide

if you are using nvidia and do not to follow the guide, you can try to use version called forge
https://github.com/lllyasviel/stable-diffusion-webui-forge?tab=readme-ov-file#installing-forge
you have >>> Click Here to Download One-Click Package<<< link there
that is for nvidia only

[ornate elk]

Nope sry, Do you have an nvidia gpu ?
Then you can take a look at the install guide from Aitrepreneur on Yt

[ornate elk]

Why should it download a model?

[karmic crown]

a1111 download sd 1.5 base by default if you do not have any models

he probably talks about that

[ornate elk]

He already used it before so I'm confused xD

[cosmic furnace]

Usually if its not there. he download it ? well this 1.5 default - v1-5-pruned.safetensors

[ornate elk]

For me it didn't downloaded it

I had to move a model over

[cosmic furnace]

Why amd adrenaline shows as if max gpu power 300 ? thats lil bit not like that)

ive also did 2-3 3dmark benchmark runs and everything was fine...but i guess sd might be more demanding -_-

[ornate elk]

Maybe run Cyberpunk at 4k while doing timespy extreme xD

[cosmic furnace]

nonono 😄

i dont need

https://tenor.com/view/explosion-mushroom-cloud-atomic-bomb-bomb-boom-gif-4464831

just gonna try not be too much sadge & do more test-settings-adjustings

[ornate elk]

Yea, but you still should consider getting a higher wattage PSU

[cosmic furnace]

Too bad you cant say to sd to not use too much powa... or too much high mhz... or overall full load on gpu

u cant? or ...?)

Yeah but itl take time =\

[ornate elk]

I dont think you can :/

[cosmic furnace]

btw u need only one or better to have both of theese?

[ornate elk]

Only the emaonly .safetensor but you can move any model over. It doesn't need to be the base 1.5

[cosmic furnace]

wuuuut o.0

emaonly is cut version on 3gb y?

i do like some additional things in nextsd.like alot of styles

on usual automatic its all a lil bit too much complicated

espec when by author name & all that

[crude trellis]

hey

I use a amd gpu and I have some doubts regarding it with stable diffusion

I've heard that if you're using the directml version of stable diffusion with amd cards, you can generate images with the same quality of nvidia

is that true? or are amd images worse at generation quality?

what is exactly the difference, performance? loading time? generation quality?

thanks in advance for whoever answer this

[cosmic furnace]

only speed & amd catch up) . btw even without zluda turbo models r pretty fast

1280-1024 text2img generation no problems o.0

is this still slow?

[ornate elk]

Hey, you can generate the same quality on each device, being amd, nvidia, Intel or CPU.
The difference is mostly in the performance.
AMD with SD works the best on Linux right now.
For windows you had the Directml webui fork of auto1111. Which is good in support and usability but lacks the performance and vram optimisations.
But since 2 weeks we also have the Zluda Backend which is much faster and less vram limited.

[crude trellis]

thanks

[cosmic furnace]

If it can generate high resolution when txt2img not img2img than it means issue is smwhere else?

[crude trellis]

so technically I can replicate any image from civitai or hugging face with AMD as long as I can set up the right prompts, right?

[ornate elk]

Yes, but due the hardware difference you can't get the exact same image like an nvidia user. But for example I can share an image from my amd SD instance and you could recreate 1to1

When using the same settings and models etc

[cosmic furnace]

erm

[ornate elk]

2its is not that fast

[cosmic furnace]

it used all 20gb vram with zluda for 1408-1152 ?

[crude trellis]

I have a RX 6750XT with 12gb VRAM

that should be enough for SD, right?

[cosmic furnace]

maybe i need to do some settings tweaking? ive also made downclocking)

maybe another optimizing core?

[ornate elk]

Yes I tested SD on a 6700xt a few months ago too

[ornate elk]

I can test my settings tomorrow so we can compare

[cosmic furnace]

It means on this model 1408-1152 is the limit for me? or it just uses all vram it can on zluda too ?

[ornate elk]

Depends on the settings

[cosmic furnace]

why now ive got no rebotts on high resolution? 🤔 made a clean install of sd next

also all that insta resets was mostly on img2img

img2img on sd.next not working for me? its kinda complicated and not as easy to use as directml 😦

[ornate elk]

@crude trellis if you want to start and try out the Directml fork of auto1111. (Stable diffusion)
Checkout my install guide in the Pinned Messages of this channel.
The newer guide is for the Zluda Backend its Experimental so please Start with the other guide. (Not zluda)

[cosmic furnace]

i cant quite catch where png info tab is =\

[ornate elk]

Me to the SD.Next ui is confusing

[cosmic furnace]

just how much slow is this?)

40 steps 1408-1152 in 20 sec still slow? 🤔

[crude trellis]

I already have stable diffusion directml, but I've been facing some problems

[ornate elk]

Oh okay, and which problems?

[crude trellis]

it's not happening now, but a while ago I couldn't generate any images because of some vram error

there was some messages saying that I haven't enough vram

but like I have 12gb and there are people who can handle SD with 4gb

is that linked to how heavy the settings are?

[ornate elk]

Ah the basic error for amd.
Make sure you have this in your webui-user.bat:
--medvram --opt-sub-quad-attention --opt-split-attention --upcast-sampling --no-half-vae

After you edited it save and then read my AMD quick info and guide for upscaling.
Here: (its a comment in an issue thread)
https://github.com/lshqqytiger/stable-diffusion-webui-directml/issues/373#issuecomment-1936982157

[cosmic furnace]

can it be that img2img not working for me? hmmm gotta try 🤔 . img2img not using more gpu res than txt2img right?

[ornate elk]

I will test it tomorrow

[crude trellis]

okkk

I have another problem tho

[cosmic furnace]

it cant start

always this error o.0

[ornate elk]

Img2img?

[crude trellis]

its about the results of generation so I guess it can be more complex to describe idk

but basically all images that I tried to gen are all kinda blury and "dead", I can't get the same results as when I was using SD on google collab

even if i grab a ready to go prompt on civitai and just paste it on SD, my images are always worse

and yes I am using the same models

[cosmic furnace]

yes

[crude trellis]

this one, for example

[ornate elk]

For that you need to show us an example but blurry, seems like youe using Hires fix with latent upscale with 0.7 denois which isnt good

[cosmic furnace]

it was 540 error now 590

[crude trellis]

or this one

[cosmic furnace]

dyu know why? dang

[ornate elk]

Ah okay, thats not "using the same settings" mostly people use hires fix and a VAE file for color correction

[ornate elk]

You definitly need a VAE file

Kl-f8-anime2 would be good

[crude trellis]

what is this?

[crude trellis]

most recent one

you see, all of them are kinda blury

[cosmic furnace]

y need sharpness... how much steps?)

[crude trellis]

20

[crude trellis]

hmmm

[cosmic furnace]

u shud look for recomended sampling-steps-cfg

for this exact model

some models need more & vae -other files & presets

[ornate elk]

Thats an example, the VAE is for color fix

[cosmic furnace]

it refuses to do img2img for me at all o.0

even to lower resolution than i did with txt2img

am i doing smthing wrong?

resize method non or some of this?

or u not familiar with next sd? 🙃

[crude trellis]

I wasnt really trying to gen seriously, all of those images were just tests

but one time I tried to copy as nearly as possible some prompts and settings from civitai without changing anything just to see if it was working, but my images were always worse... thats why I tought it could be AMD related

og: https://civitai.com/images/394741
my result:

you see what im talking about?

I hope this VAE thing helps

[cosmic furnace]

same errors same vram usage when lower resolution...

Why image hidden its not nsfw 😏

[crude trellis]

idk, some people may consider it nsfw

[ornate elk]

Its not an AMD thing. Its just a missing vae file.
Thats normal for anime models.

Here is a copy of my tip for a user that had the exact same problem 2 weeks ago.

Hey, first you need a VAE file for better colors.
To activate the VAE menu go into
Settings, User Interface, Quicksettings, there add sd_vae
Then hit apply and reload ui. Then you can download a VAE (like kl-f8-anime2, 840000-mse, or clearvae) and put it into models/VAE folder And select it in the new dropdown.

The actual blurry edging comes from the wrong settings of your Hires Fix.
You have to set the hires steps to 10.
Then choose Resrgan4x Anime6b for example as upscaler.
The Denois should be 0.5 or lower.

For latent upscalers try bicubic and stay at 0.5 denois

[crude trellis]

okei friend

[crude trellis]

btw, heres the message I was talking about

going to do what you told me now to see if it works

[cosmic furnace]

sdnext discord seems to be sleeping 😦

its a lil bit sadge that i cant use img2img =\

[ornate elk]

Do the VAE settings from my last message. Then check that you have the same webui-user.bat commands as I mentioned. Then it should work much better. And for upscaling (what improves quality) you can read the linked comment from me, with tips.

I'm going to sleep now. Its pretty late here. Feel free to @ me if it worked or if you have more questions.

[crude trellis]

okay, thank you

I've asked the same questions in about 2 or 3 servers, this one is the only one I was answered

[ornate elk]

Btw, i made that on a 6700xt, a few months ago.