🤝｜tech-support

[remote agate]

I copied the parameters directly from PNG info, same seed and same prompts and same configuration

[ornate elk]

can you show me the png info?

and now drop the other image in png info

[remote agate]

what image?

[ornate elk]

the failed one

[lost crescent]

AUTOMATIC1111 takes soooooo loooooong to make YAML files - but why? 🙂

[ornate elk]

can you show an example? cmd log?

[vast crest]

Hi everyone, I have a problem with a checkpoint not working for me. I download it and insert it into the folder: Models then into the folder: Stable diffusion. I try to load the model but nothing happens. Is there anyone who could help me?

p.s. Are there Italians?

[ornate elk]

hey, make sure the model is larger than 1.98gb

you see the difference?

[vast crest]

is 6,4 GB

[ornate elk]

ok whats your GPU ?

[vast crest]

but I also have 7.5 Gb models but they work for me

My Gpu: RTX 4070 OC 12 GB

[remote agate]

There's no difference

[ornate elk]

there is, look closely, in your first image you used something strange called "cfg rescale phi:0" i ve never seen before. so maybe from an extension

[ornate elk]

okay, is your webui updated?

[vast crest]

I'm going to check now

[ornate elk]

you can check the versions at the bottom of the webui in browser

[little echo]

so is it possible to open up an SF model to see what tags it uses?

[ornate elk]

i dont think so

[vast crest]

I don't see anything. I downloaded this model from Tensor.art. Can I send you the link so you can check it too just to be sure?

[ornate elk]

@little echo and to answer your other question, a model will always change the output image, even when you type in words that doesnt exist

[remote agate]

It is a parameter of an extension, but it is disabled. so no, there is no difference, I know because with "--no-half" I get almost identical results, with very small differences.

[ornate elk]

maybe try enable the extension again set the value to 0 and recreate?

to check

[remote agate]

I think perhaps the problem lies in the fact that the graphic card is not capable of correctly optimizing the processes, so that's why I get very deterministic results with "--no-half"

[ornate elk]

i dont think so, you can get nearly the same result with any card

[vast crest]

I use these commands in the webui: --autolaunch --xformers --opt-sdp-attention --precision full --no-half --no-half-vae --disable-safe-unpickle

[ornate elk]

xformers will slightly alter the details, but not to deformation

[vast crest]

However, this is the model that doesn't work for me. my stable diffusion can't load it on the browser page: https://tensor.art/models/647319333476953074

[ornate elk]

wow that slows down a lot and with that sdxl wont work
remove --precision full --no-half --disable-safe-unpickle --opt-sdp-attention
that will speed up your webui a lot

[heady sorrel]

Hi guys! I'm trying to do img2img with the reactor extension activated, but when i sample with some SDE sampler i receive some errors

UnboundLocalError: local variable 'h' referenced before assignment

Only with SDE Sampler

So DPM +2 SDE Karras

DPM +3 SDE Karras

with others samplers works ok

[vast crest]

I'll be back later guys, I have to go eat pizza 🙂

[remote agate]

It may be an error when manufacturing my graphics card :/

[umbral plover]

What does this --disable-safe-unpickle do?

[heady sorrel]

After click generate

without reactor extension works fine the SDE Sampler

[little echo]

are the tags in the prompt that are recognized contingent on only the selected SF model at a time? or is there like an underlying model that the models build upon

[heady sorrel]

You asked about my error?

[covert fox]

I reinstalled SD and use same prompt, same LoRa and same parameter but the result was not same, How can I fix this?

[ornate elk]

it makes you vulnerable to malicious files (old models)

[ornate elk]

nah i dont think so, try to create a normal image, and then restart the pc and try recreate it, im sure it will be the same or nearly the same

[ornate elk]

okay, let me know if it worked

[remote agate]

Furthermore, I tested the parameters of my image from Google Colab Pro, and the results are entirely replicable without the "CFG rescale" extension.

[little echo]

no i am just asking about how SF tags work

are the recognition of tags based only on the current SF model?

[remote agate]

Yesterday I tried to do that without the "--no-half" option, and I encountered errors after attempting to recreate the image again without changing the image parameters.

[little echo]

do you know @ornate elk

also no need to respond immediately if u busy

[remote agate]

That is to say, initially, I obtained an almost identical image, but after attempting to recreate it again, errors emerged.

[ornate elk]

try recreate a new image, not that one with the extension meta data

[half island]

Hoi, with comfyUI, can it read from lora if fetched like lora's in auto as just a command in positive prompts? Or does it need the lora node as well?

[ornate elk]

i cant really explain how Stable-Diffusion works. Its pretty technical. You should look up the Original documents of how the dataset is created, but let me say the models were trained on images, these images got tagged with words, and these words will be recognized when generating

[ornate elk]

check the meta data in PNG info fro mthe old and the new one and compare them

[ornate elk]

is your webui updated?

[remote agate]

Without changing any parameter and without extensions.

[ornate elk]

so it works

[remote agate]

And as I mentioned earlier, if I keep trying, it will start to distort the image.

[remote agate]

not really

[ornate elk]

then i would suggest deleting the ui-config.json and the venv folder. that will make sure your on a clean installation

then there shouldnt be any hidden stuff thats messing up in the background

[remote agate]

I did it, I started the web UI, and it began to install this:

[ornate elk]

thats okay

its stuff from the venv that gets rebuild

[remote agate]

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.

[remote agate]

The images no longer deform, but they do create significant changes if I try again

[ornate elk]

can you show me the settings you used when getting this error?

[remote agate]

webui-user.bat or image settings?

[ornate elk]

txt2img settings

resolution, steps, etc

[remote agate]

Literally the same configurations of this image (without the extension)

[ornate elk]

but thats not the error you got

[remote agate]

appeared that error and did not generate any image

[ornate elk]

you got that error just by starting the webui?

[remote agate]

I clicked on "generate" and got that error; it rarely happens, but I don't know how to replicate it because I just click "generate" again with the same settings, and the error doesn't appear again

[ornate elk]

okay thats very strange

[remote agate]

I don't mind that error; what does concern me, though, is that I may not be able to recreate images without using "--no-hafl"

[remote agate]

I followed the steps you instructed me to delete the files, and it worked temporarily, meaning I was able to recreate the same image 5 times. However, upon restarting the page, I started experiencing errors again :(

[covert fox]

The Model is a different location

[ornate elk]

from when is the older image?

the model name isdifferent

[vast crest]

Hi, I'm back, I made the changes you told me to do.
the model works for me, but the images I create come out like this

Then here are all the Versions: version: 1.7.0  •  python: 3.10.11  •  torch: 2.0.1+cu118  •  xformers: 0.0.20  •  gradio: 3.41.2  •  checkpoint: 71bcd704cb

[ornate elk]

can you show the txt2img settings?

your vae is mostly incorrect

sdxl needs a sdxl vae

and 1.5 needs 1.5 vae

[ornate elk]

wrong vae

sdxl need an sdxl vae

orange vae is 1.5

[covert fox]

older was created on 08/10/2023
newest is created today
the model name is different, yes but I copied it from an older file before deleting SD

[vast crest]

where can I find the missing files I need?

[ornate elk]

here:
https://huggingface.co/stabilityai/sdxl-vae/tree/main

[vast crest]

ok thx

[ornate elk]

did you changed GPU ?

between then and now?

[covert fox]

No, is both gen with SD on Google Colab

[vast crest]

it finally works, thanks a lot 🙂

[ornate elk]

ah okay, then its collabs fault and they updated something that changed it mostly

[ornate elk]

no problem 🙂

[covert fox]

How to find a way to fix it?

[ornate elk]

you can go into settings and check that this is set to CPU:

or NV if collab only uses Nvidia

[remote agate]

@ornate elk I don't want to sound insistent, but I believe the problem might not be with Automatic 1111; rather, it could be within my PC. What could be causing this issue? I can try various solutions.

[ornate elk]

i can give you an image to recreate

[vast crest]

I would also have a problem after I install TensorRT. It doesn't give me any errors, but nothing appears on the main stable diffusion page. I find it in the folder: extensions but otherwise nothing to do. How many problems I have with SD XDDD

[ornate elk]

if you dont want problems, dont use TensorRT xD

after deleting the extension you also need to delete the venv folder

[remote agate]

okey, preferably it should be a 1.5 SD model please

[vast crest]

Already done calm down 🙂

I saw many videos on YouTube that spoke well of tensorRT and I wanted to try it, but nothing happened. In the end, what is it really for? JUST to improve the images?

[zealous cargo]

why is sd giving me random images?

[ornate elk]

TensorRT enables the Tensor cores (the raytracing cores) to work for image generation. That makes the process faster. But it isnt compatible with every feature of SD

[ornate elk]

cn you explain a bit more?

[zealous cargo]

it ignores my promts and gives me weird pictures with deformation

[ornate elk]

can you show an example?

[ornate elk]

sure, here i used dreamshaper v8:
cute penguin, masterpiece, best quality, highres, detailed, Negative prompt: blurry, deformed, Steps: 30, Sampler: Euler a, CFG scale: 7, Seed: 505050505, Size: 512x768, Model hash: 879db523c3, Model: dreamshaper_8, VAE hash: 235745af8d, VAE: vae-ft-mse-840000-ema-pruned.ckpt, Denoising strength: 0.5, Hires upscale: 2, Hires steps: 10, Hires upscaler: R-ESRGAN 4x+, Version: v1.7.0

[remote agate]

Could Nvidia drivers cause incorrect calculations in fp16? 🤔

[ornate elk]

i dont think so, maybe if they are really really old

make sure to be on the latest update

[vast crest]

Thanks so much for the explanation, and thanks again for the help 🙂

[ornate elk]

no problem 🙂

[snow lance]

fresh install ubuntu 22.04 with python.3.10, keep it or add 3.11/3.12?

[ornate elk]

SD only supports 3.10.6 up to 3.10.13

[snow lance]

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu121

instead of this one?

pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu121```

[ornate elk]

no need for nightly

[little echo]

ok but can you atleast confirm if this is correct:
the SF models can define their own tags
at the same time how some of the tags are used/identified is beyond just the specific model being used?

[ornate elk]

no the models cant define their own tags, the images of the database the model got trained on got tagged

people can train their own model by tagging images and then running a script to create the model and it will recognize the new words used inside the tagged images + the already known words from the base model that is needed for training, (so you can teach a mode lhow something looks like or is named)

[little echo]

ok thanks i will delve more into it if i have a good moment

[ornate elk]

no problem 🙂 yea take a look at documents from stability or runway like this one:
https://huggingface.co/blog/stable_diffusion#how-does-stable-diffusion-work

[cloud topaz]

hello! i use comfy ui with reactor fast face swap, and the result image has always a low quality on the face, any idea?
thanks!

[glad fiber]

can someone help me here please #🤝｜tech-support message. Been stuck on that for some days

[ornate elk]

Because reactor normaly applies facesrestore via codeformer after the swap

So you would need to include that in the workflow

Ah I see you only need to select a face restore model in the reactor node

Choose codeformer

[cloud topaz]

okay thank you works now 🙂

hey reactor face swap is working great but the face is a little too smooth, any fix?

[ornate elk]

You can lower the face restore visibility or the codeformer weight

[cloud topaz]

thanks

i lowered the codeformer weight to 0

as you can see the face is very smooth

[zealous cargo]

is it possible to add something like -xformers to comfyui?

[cloud topaz]

and the face im providing isnt so smooth

[ember venture]

I'm getting crashes in trying Dreambooth training in SDNext, can anyone help?

[remote agate]

with "--no-half"

[rustic echo]

I’m in the process of getting a 3d model made for my character so that it can be trained on a LORA… The big question I have is
If the model I get back has a plastic/animated look to it (such as some of the examples below)
If I use a photorealistic checkpoint on the trained LORA…
Will I have a believable photorealistic character?
(I’m trying to gauge how stressed I should be on hiring the best 3d modeler that will take care of texture/skin)

[steady thistle]

I still can't get Inpaint Anything to not fail on the first step. Anybody else using AMD but who got it working?

On "Run Segment Anything" it fails immediately. Inpaint Anything - ERROR - The size of tensor a (0) must match the size of tensor b (256) at non-singleton dimension 1

Okay! I searched on a forum and found that there's a "Run Segment Anything on CPU" in Settings. That did work for me. Slow as Christmas but it is still easier than masking by hand.

[steady thistle]

But now it won't actually inpaint or do anything once I got the image masked. RuntimeError: mat1 and mat2 must have the same dtype

[snow lance]

linux / comfyui users, what do you guys have setup for swap? 64g ram

[ornate elk]

Can you upload the original? So I can check the meta data

[ornate elk]

You mean the freckles?

[remote agate]

i use the same settings from your image

[ornate elk]

Have you put the PNG in info tab and selected send to txt2img?

[remote agate]

Yes

[steep belfry]

When Stable Diffusion goes into production, it looks like this. I still don't know how to make images that are generated for 50 seconds faster, and so...

Ah, the other day, I was suggested to read a guide, but I couldn't find it

[ornate elk]

What's your GPU and what's inside your webui-user.bat?

[steep belfry]

GPU:NVIDIA GeForce GTX 1660
Webui:

[ornate elk]

What tool or webui are you using?

If your using auto1111 pls show the content of your webui-user.bat

Not webui.bat

[steep belfry]

@echo off

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

call webui.bat

#!/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"

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

[ornate elk]

In the webui-user.bat
At the line COMMANDLINE_ARGS=
you need to add: --xformers --medvram --no-half
Then save and relaunch
It should be much faster then

[remote agate]

Well, things have gotten worse; now it's unable to display images correctly even with "no-half," and I truly feel very confused

[ornate elk]

Can you show an example?

With the txt2img settings

[remote agate]

Literraly is just noise

[ornate elk]

Which model?

[remote agate]

any model

[ornate elk]

Which settings?

Is your GPU driver updated?

[remote agate]

I checked it, it's updated to the latest version

[remote agate]

I tried with the same settings that I had shared with you a few hours ago.

[steep belfry]

continues the same way

[ornate elk]

That means you didn't edited the webui-user.bat correctly or didn't launched the webui-user.bat

[steep belfry]

Did I put it in the wrong one?

[vocal burrow]

no but you probably launched webui.bat instead of webui-user.bat

[vocal burrow]

What do where python, python --version and py --list-paths say ? (run those in cmd)

[steep belfry]

python: in SSD
Version 3.10.6

[spice harness]

Is anyone from the stability AI team available for a quick chat? I have a question about hosting an event where we will rely on stable diffusion, and i would like the process to run as smoothly as possible

[steep belfry]

another time I continue, I'm sleepy and then I'll come back

[ornate elk]

You can try to delete the venv folder and relaunch the webui-user.bat

Feel free to ask here again later

[grand fable]

try to add something to the prompt.. something like "skin imperfections" "freckles" .. or "smooth skin" to the negative prompt..

[cloud topaz]

Okay thank you I will try but you think it will work even if I’m using face swap?

[wraith skiff]

i cant upscale

with GFPGAN

here cmd

[grand fable]

don't know, just a tip.. but please let me know if it works and send the final image 😉

[cloud topaz]

Okay I will show you this evening when I get home 🙂

[novel spade]

I deleted my automatic1111 stable diffusion folder by accident and i believe i broke it upon recovery. im getting alot of traceback calls and im pretty sure its an issue with spare characters that shouldnt be in the encryption

i deleted VENV and ran webui and hten WebUI_user to rebuild VENV but i get the same errors

this is my current error log, i believe there is a lot less htan before rebuilding venv

would i just be able to recompile with git to fix this? (im a noob with no coding experience btw)

[ornate elk]

Can you show the settings you used?

[ornate elk]

Hey try to delete the config.json

And the venv

[novel spade]

gotcha

[ornate elk]

Also add
Git pull
To your webui-user.bat

Before starting it again

[novel spade]

i did just try that - ill delete venv again, delete config.json, then relaunch user.bat

git pull is added to the bat

[wraith skiff]

thats all

[novel spade]

seems to be an error when using git pull

[ornate elk]

okay, then its maybe missing the GFPGAN model.
you can download this one and place it inside the stable-diffusion-webui folder.
Then relaunch the webui-user.bat

[wraith skiff]

im tryin

[novel spade]

that definitely did something, i got way more errors now :S

[ornate elk]

@novel spade best would be a quick reinstall and then moving the models over

[novel spade]

thats what i was thinking - i wasnt sure the best way to go about that

[ornate elk]

check my install guide in the Pinned Message of this channel. You can set it up in less than 10 minutes

[wraith skiff]

still same, i put in C:\all\stable-diffusion-webui

[steep belfry]

it got better, thank you

[novel spade]

i can safely move over things like textual inversions/loras as well?

[ornate elk]

yes

[novel spade]

awesome, thanks CS

[ornate elk]

np, just dont copy the venv or extensions over, only models, loras, embeddings, upscalers, outputs

[novel spade]

ah, i will have to manually reinstall extensions?

[ornate elk]

via the extensions tab inside the webui

[novel spade]

okay no problem

[ornate elk]

controlnet models can be copied too

if you used them

[novel spade]

awesome, i did yeah

ill copy one large section at a time and test it so i can see if something goes wrong

[ornate elk]

if you copy one model insidethe new models/stable-diffusion folder, the initial install will be faster. (because a fresh install would download the 1.5 base model)

[wraith skiff]

[ornate elk]

then show the me the settings and there search for face

[ornate elk]

can you try an other upscaler ?

[wraith skiff]

i only try this one

any suggestion?

[novel spade]

sounds good

[ornate elk]

to test, try Resrgan4x+

[wraith skiff]

generate

🙏

ai god help us

nah same

[novel spade]

is sd okay to install on an external drive?

[wraith skiff]

i tried swinir and gfpgan off its worked

but when i on gfpgan

i guess problem is gfpgan not upscalers

[ornate elk]

you can yes. its better to install SD on an SSD. But you can store models on external hdd to not fill your C drive for example

[granite quest]

hi guys, can i get the version of sd A111 that was used to generate an image? i upgraded to v1.7.0 to try SDXL but i want to revert back to the previous version (which i forgot)

[ornate elk]

do you know which version you used before?

and hey

[granite quest]

i think it was 1.3.x but i forgot the last number

[ornate elk]

oh thats and much older version and not good supported anymore

if you really want to get it you would need to know the commit hash of that version and then git clone it new from the github repo

[granite quest]

ohhh thats unfortunate, okay i guess i'll stay at 1.7

thanks g as always 👍

[wraith skiff]

[ornate elk]

np, what is missing in 1.7 from 1.3?

[granite quest]

i got this error: https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/8084
but i just made the fix as said by the comment there, im worried that the quality of the images i have will be affected because i these command line args now:
--no-half-vae --disable-nan-check --xformers --medvram --disable-model-loading-ram-optimization --precision full --opt-split-attention --no-half

[karmic crown]

do not use --disable-nan-check

[ornate elk]

whats your GPU?

[ornate elk]

can yo show me the settings of face restore?

not the extras tab

[granite quest]

its a GTX 1660 Ti, 6gb of VRAM

[wraith skiff]

where is face restore settings ? 🤷‍♀️

[granite quest]

i think i made this option because for some reason SD just produced black images iirc

[ornate elk]

in the settings under Face Restoration

[ornate elk]

remove everything and only use --medvram --xformers --no-half

that wont change quality, but makes everything faster

[wraith skiff]

no need

i guess i fixed it

[wraith skiff]

i put in models/gfpgan

thanks for help 👍

[ornate elk]

ah okay, good

np

[granite quest]

alright, ill report if anything arises

thanks again man 👍

[ornate elk]

yes, no problem 🙂

[novel spade]

trying to add ownership to new directory using " git config --global --add safe.directory path" and its telling me fatal: not in a git directory

i forget what im doing wrong

am i not supposed to use this command in git bash?

[ornate elk]

this is needed when your external drive isnt formatted correctly

it would need to be ntfs

[ornate elk]

you would need to run that in a cmd inside the folder

[novel spade]

it wont work in exFAT at all?

[ornate elk]

i cant guarantee that it would work or wont cause problems later

the command should work, you may need to use it mutliple times for different folders

[novel spade]

it just tells me fatal: not a git directrory

[ornate elk]

you need to run the cmd inside the stable-diffusion-webui folder

[novel spade]

mhm thats what im trying

[ornate elk]

okay, then no luck

i think exfat doesnt keep ownership at all

[novel spade]

ah okay

no problem

ill try on another drive

[novel spade]

that worked, while moving my embeddings it seems to think one is possibly malicious, is there any way to make an exception?

deleting the .pt file for it stops the error so its the obvious culprit

[ornate elk]

you can redownload it

[novel spade]

word

[ornate elk]

if you find it

.safetensors are the new safer files
so stick to them if you can

[valid gorge]

Anyone tested Intel ARCs, are they any good with SD?

[steady thistle]

@ornate elk I got Inpaint Anything to finally start, but I can't actually process anything. I keep getting this error. RuntimeError: mat1 and mat2 must have the same dtype

[karmic crown]

check dimensions you are trying to use, that usually happens with wrong dimensions

[heady sorrel]

Yess

https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/13761 is same error

but nobody explained how to fix

[earnest basalt]

Not sure if this is right tab. Could someone assist or point me in direction of assistance to install reactor correctly. I make a lot of signatures using face swap and I previously used roop. I'm now wanting to make gif or small video signatures and read reactor is the way to go. The issue I'm encountering is that when I faceswap using reactor almost nothing happens to original image bar the upscaling.

So far I've tried.

Disabling roop
Renaming roop folder in the webui to insightface
Installed force reactor with no change either

Any help would be greatly appreciated, thanks 🙏👍

[karmic crown]

"issue seems to be at all samplers from DPM++ 2M SDE and DPM++ 3M SDE
upping steps to over 100 also works as workaround."

so, lower the number of steps or change the sampler

[languid ledge]

it works now but uhh are you familiar with reactor cus something doesn't looks right

[steady thistle]

You mean the size of the image? Does it need to be divisible by 16 or something?

[karmic crown]

yes

[steady thistle]

Ah, okay, will try that.

[ornate elk]

Sure, what's the issue?

[ornate elk]

Hey,
delete roop from the extensions folder.
Then delete the venv folder.
Then launch the webui-user.bat
And follow these install steps carefully:
https://github.com/Gourieff/sd-webui-reactor#installation

[earnest basalt]

Appreciated. I thought it would entail completely deleting roop, the initial tutorial I watched said just to disable it. I'll try that now. Thanks 👍

[ornate elk]

Let me know if it worked 🙂

[steady thistle]

I tried an image at 1024x1024 and I got the same error. It is on an AMD. I got Inpaint Anything purely by having it run on CPU in the settings.

[earnest basalt]

Just wanting to make sure I do this right the first time. I've removed all of roop and reactor to re install. I cannot seem to locate the venv folder though

[karmic crown]

it is inside your stable-difusion-webui folder

[earnest basalt]

Got it. Was hidden for some reason

Just outright delete it yeah?

[karmic crown]

you can safly remove it, it will be created again when you run webui-user.bat

[steady thistle]

Sorry to bug you, but using 1024x1024 I'm just still getting RuntimeError: mat1 and mat2 must have the same dtype

[karmic crown]

can you show me the whole error log?

in many cases actuall error ocurs much earlier, and the last error is just a consequence of original error

[steady thistle]

Oh, sure. (just tried 512x512 and no luck) but here:

[earnest basalt]

When I re ran bat it installed everything as normal then closed after I pressed any key.

Then I reran and it's been stuck on installing requirements for around 5 minutes without doing anything. It's been so long since I done this process. Is that normal

[karmic crown]

i want to see if there is any other error before this one

[steady thistle]

Total progress: 100%|████████████████████████████████████████████████████████████████████| 8/8 [02:21<00:00, 17.73s/it]
2024-01-16 10:20:32,797 - Inpaint Anything - INFO - input_image: (512, 512, 3) uint8█████| 8/8 [02:21<00:00, 12.56s/it]
2024-01-16 10:20:33,095 - Inpaint Anything - INFO - SAM is running on CPU... (the option has been checked)
2024-01-16 10:20:33,098 - Inpaint Anything - INFO - FastSamAutomaticMaskGenerator FastSAM-x.pt

0: 512x512 34 objects, 1109.3ms
Speed: 9.0ms preprocess, 1109.3ms inference, 66.9ms postprocess per image at shape (1, 3, 512, 512)
2024-01-16 10:20:35,391 - Inpaint Anything - INFO - sam_masks: 34
Processing segments: 100%|████████████████████████████████████████████████████████████| 34/34 [00:00<00:00, 121.89it/s]
2024-01-16 10:20:59,795 - Inpaint Anything - INFO - Loading model runwayml/stable-diffusion-inpainting
2024-01-16 10:20:59,798 - Inpaint Anything - INFO - local_files_only: True
2024-01-16 10:21:14,127 - Inpaint Anything - INFO - Using sampler DDIM
2024-01-16 10:21:14,610 - Inpaint Anything - INFO - Enable attention slicing

[karmic crown]

just leave it as it is, it might take some time depending on your internet and computer speed

[steady thistle]

That's what was just before the traceback, that was going up to the image production I think. Let me double check.

Yes, previous to that was building the 512x512. I'll restart it and go straight to the Inpaint Anything if that's best?

[karmic crown]

make sense

"It likely means you are trying to use LoRAs incompatible with the currently loaded checkpoint, like using 1.5 LoRA on 2.1 or SDXL checkpoint and vice versa. Make sure you have matching pairs."

check this

can you try with just a model, without any lora or external vae?

[meager carbon]

@ornate elk one last question... Really sorry about it but, what part of the SDXL model do I put in the models folder? The whole thing?

[karmic crown]

what is the "other part"?

[meager carbon]

this

[steady thistle]

Okay, here's from the beginning of running SD to the error.

[karmic crown]

just download sa

[meager carbon]

I was provided a seperate VAE

[karmic crown]

first one goes to models/Stable-diffusion

[meager carbon]

#💬｜general-chat message

[karmic crown]

and vae goes to models/VAE

[meager carbon]

ah, I see. thank you

greatly appreciated

[karmic crown]

see this

Ok so, i managed to solve it, there is a setting that says:

Lora/Networks: use old method that takes longer when you have multiple Loras active and produces same results as kohya-ss/sd-webui-additional-networks extension

in the settings ( settings -> show all -> find "LORA" string for quick getting there )

if this is enabled, it causes this issue and wont allow you to use Loras, just disable it and it should work ( it worked for me )

[steady thistle]

Oh wow, okay, I'll try that. Thank you very much

[earnest basalt]

Thanks , the whole deleting the lot worked. It did however take a lot longer to faceswap. Will this quicken over time.

With roop it took around 3-5 seconds per image.

It took around 30 seconds to do one faceswap there with same settings.

Denoising 0
Euler a
35 sampling
Image size to scale

[karmic crown]

i am not sure what is the real cause of your error, so i am searching for answers

[steady thistle]

It was not checked in settings. I'll toggle it to test.

[earnest basalt]

It's ok for speed. Just now and again it seems to take a while

[meager carbon]

do I select the XL as a checkpoint?

[karmic crown]

and use 1024x1024, 768x1024, 1024x768 or 768x768

[meager carbon]

nono I get that

like how do i set it as a base

[steady thistle]

No luck on that.

[meager carbon]

so this was a yes?

[karmic crown]

yes

[meager carbon]

thanks

[karmic crown]

models trained by users are much better than default one

[meager carbon]

figured

[karmic crown]

you can find them on https://civitai.com/

[meager carbon]

ones that have XL in the name are presumably... SDXL?

[karmic crown]

https://civitai.com/models/119012/bluepencil-xl this one is my favorite anime model

yes XL = SDXL

[meager carbon]

haha

that's the one i already had

[meager carbon]

oh I see

[karmic crown]

you always have info here

what was the base model

just do not mix xl and 1.5 resources

like sdxl model and 1.5 lora

[meager carbon]

the main reason I wanted to set up SDXL is because people generally get good looking hands out of it atleast from what I can tell...

unfortunetely that is not the case for yours truly

[karmic crown]

https://civitai.com/models/241445?modelVersionId=272411

try this one

in most cases hands are fine

[meager carbon]

like on all the bluepencil ones, the gallary images, all of them have great looking hands

[karmic crown]

you can try this negative

deformed, blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, no color, weird colors, censored, deformed glasses, lowres, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, (((worst quality))), (((low quality))), (((normal quality))), jpeg artifacts, signature, watermark, username, low resolution, bad anatomy, siamese twins, (worse quality, lower quality:1.5),

[meager carbon]

Thank you.

[karmic crown]

or this one

[deformed | disfigured], poorly drawn, [bad : wrong] anatomy, (([extra | missing | floating | disconnected | fused | malformed | mutated | abnormal | huge | disappearing] [limb | arm | leg | head | hands | fingers | feet | calf | thigh])), (([short | fat | ugly | crooked] fingers and hands)), ((two thumbs on one hand)), ((two nails on one finger)), four fingers, six fingers, long neck, blurry, ugly, deformed, noisy, low poly

but CS1o told me that he had some problems with this one (that works fine for me)

[steady thistle]

I very much appreciate your time, I can't think how to move forward on Inpaint Anything. Can you recommend a different extension that helps with inpainting?

[karmic crown]

i havent use inpainting yet, to be honest

[steady thistle]

Ah, makes sense.

[meager carbon]

using this gave me this:

haha

[karmic crown]

are you using 512x512 with blue pencil XL?

that is no go

[meager carbon]

1024x1024

[karmic crown]

hmm

show me the info below the image

sampler, steps, dimmensions...

[meager carbon]

I already generated another, but sure one sec

Steps: 20, Sampler: Euler, CFG scale: 7, Seed: 2773488136, Size: 1024x1024, Model hash: 77b1007aa5, Model: bluePencilXL_v310, Version: 1.7.0

[karmic crown]

do not use Euler

[meager carbon]

oh?

Alright, which one should I use then?

[karmic crown]

DPM++ 2M Karras
Euler a

for example

[meager carbon]

ah okay

[languid ledge]

could I DM you about it?

[karmic crown]

i would ask that in #📝｜prompting-help
this thread is for technical issues and problems with installation

[karmic crown]

is there any difference?

[meager carbon]

slight.

[karmic crown]

are you using one specific seed, or you have -1?

[meager carbon]

euler A

[meager carbon]

-1

I guess this is promting help at this point

thanks for your help

[little bough]

Hi, is there any known issue with the latest version of controlnet? I installed the faceid ip adapter this morning and decided to update all my extensions, controlnet etc, tried to use faceid and getting CUDA out of VRAM errors and not just with that but regular ipadapter.. everything was working fine right before i updated... there was another update a little while ago since this morning and I've updated but the problem is still there, anyone else getting this? (I'm on an 8gb card

[karmic crown]

be sure to use correct ipadapter model

be sure you do not have something like this

xl model with sd15 adapter

thing like that will not work

if you are using xl mode, be sure to use this

[little bough]

im not new to ipadapter, like i mentioned, I was using ipadapter just fine without errors, after updating controlnet, im getting errors

return t.to(device, dtype if t.is_floating_point() or t.is_complex() else None, non_blocking)
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 20.00 MiB (GPU 0; 8.00 GiB total capacity; 7.04 GiB already allocated; 0 bytes free; 7.30 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[karmic crown]

i am just saying to check your settings, since i overlooked those many times

[half island]

Hoi, any idea why my comfyui suddenly can't effectively load images with generation data by drag and drop? Not even load button barely works, and sometimes i have to clear the current node order to load one, but now nothing works.

[little bough]

did something in controlnet break?

[little bough]

yeah im aware of mismatches

[karmic crown]

i have not update mine for some time, but i am on mac, so even if i update to latest, i might not experience the error

[little bough]

do you know a way to flush the vram out being held without restarting a1111?

[karmic crown]

https://github.com/Mikubill/sd-webui-controlnet/issues

i do not see any issue raised in a last few days related to your problem

[karmic crown]

restart is the only way i am aware of

[little bough]

indeed

[little bough]

ok thanks. i had recalled someone before saying switching models would do it but it definitely isnt lol

[plain axle]

Hihi! I have a 2T external harddrive coming in on Sat and would like to install stable diffusion on it.

Will this be possible?

[karmic crown]

actually, try this

but i am not sure will it help

[ornate elk]

Hey, yes thats possible, HDDs or SSDs?
Best is to install Stable-diffusion on an SSD and store models on HDD to save space

[ornate elk]

Yes

[half island]

Certainly. It will just be very slow. As hdd's usual speed is 120-180MB's. Sata ssd's are 550MB's, and on gen 3 nvme's, it's 1.2GB's, and on gen 4 nvme's, it's 2.4GB's ish

Will take around 1.5-3 min to load a 1.5 model iirc from a spinner. Took around 40 sec on my server's dual actuator (it has 2 reading heads) harddrives to load over ethernet.

[plain axle]

Samsung T7 Shield 2TB Portable SSD, USB 3.2 Gen2, Rugged, IP65 Rated, for Photographers, Content Creators and Gaming, External Solid State Drive (MU-PE2T0R/AM, 2022), Blue https://a.co/d/5wkK8VM

Its this

[half island]

Ah, then it will be fast. AS you said harddrive, which is quite different from a ssd :P So yes, that one will run and load everything fast :)

[plain axle]

I'm am brand spanking new to the IT/Tech world and would really appreciate any help on Sat when I'm ready to get everything together!

[half island]

You can even have the ssd only have models that takes the longest to load if you jump from model to model, and just have basic comfyui installed on your internal storage.

[plain axle]

Yo, we're on step 1 heh.

[half island]

I've even set to have automatic1111 to load models from comfy's location as i use comfy the most.

[half island]

What's step 1? :P

[plain axle]

The only thing I use my computer for is to play games like Remnant 2 and watch twitch.

I have zero actual computer experience so please talk to me like I'm a kindergartner on Sat lol

But yeah, knowing I have ppl I can fall back on if and when I fuck everything up is huge

[half island]

Then i shall reformulate it all :)

When you say harddrive, that usually refers to these big heavy spinners. And that arm you see, that's the limiting factor, as it can only move so fast, and only fetch one thing at a time with it's "slow arm".
Ssd's on the other hand that you have ordered on the second image has no arms, or inner/outer part of any platter, it's all chips controlled by it's own processor that fetches miles faster than any arm on a harddrive.

[plain axle]

What does ssd stand for? And when I did a search for external harddrive, the thing I bought popped up so yeah, sorry

[half island]

so a harddrive for stable diffusion can use a good few minutes to load. A regular ssd? 10 sec ish. Or on your ssd, 5 sec .

[half island]

Solid state drive. Meaning, everything is solid and doesn't move.

[plain axle]

Nice! Ty!

[half island]

And all those black chips there is the storage. And it can access all of them equally fast :D

[plain axle]

Cool cool!

[half island]

If you get on my level of data hoarding, there will be a few of them in the pc xD The top one is a harddrive i tested for something once, so it's not part of the "crew" anymore lol

[plain axle]

But weather permitting, I should have all the things on Sat afternoon/evening. Will u be around @half island

[meager carbon]

your level? pff, weak.

[plain axle]

Lol you two

[half island]

for sure, i'm home all day. So i'll help with what i can. If i don't "can", i shall google to also "can" xD. Just ping me once ready

[half island]

Fine, have my server then. 4x 18TB and my old shit nas's 8TB in parity lol

[meager carbon]

still weak

[half island]

Then show your "not weak" then lol

[plain axle]

See, I tried Google and YouTube, like hours of it, b4 coming here. But I lacked context and basic knowledge to process what I read and watched with confidence. So thank you!

https://tenor.com/view/wotsits-wotsits-giants-world-record-guinness-world-record-cheesy-wotsit-gif-18341987

@ both of u

[half island]

Aye, know the struggle. Even for simple info and knowledge, it takes me 20 times longer. Especially to memorize. So i didn't get somewhat savvy with stable diffusion and basic python commands for 6 months lol. But that's because i also have "knowledge digestion disability" if i can call it that lol

[plain axle]

I gotta take a nap after I actively try to learn something, so my brain can digest, and relearn what I just learned to make it stick, b4 I can advance. So yeah, I feel ya

[half island]

Though, i'll one up senk0 with my 16k third of a GB generation xD

[half island]

Aye. I see lots of people use fancy nodes in comfyui, and i still don't know what half of them do xD

[karmic crown]

i would need about a week to upscale that i guess 🤣

[half island]

That one ate all my 24GB video memory, 32GB ram as "emergency video memory" and some of my nvme's pagefile as emergency as well lol

I really hope nvidia or even amd can make a 32-48GB card so i can effectively do my high res stuff

[plain axle]

I know you're using English, but it's not the Kings English ,sir. Kindly remember you're communicating with someone who is ignorant of your tech language varient and educate as u go heh

[karmic crown]

i saw how a100 with 80GB work, and i am still in pain after a month 🤣

[plain axle]

https://tenor.com/view/high-five-apir-news-report-gif-24038155

We'll make a Technical Editor out of you yet!

Alright, gotta go. Randomly came across this old as heck show called Californication with David Duchovny, and kinda hooked.

[half island]

Gotchu, you can read this when you get back :P

Sorry, what i'm saying is that there's a stable diffusion generator that has so many functions i don't know what half of them do :P

What i do though as i can't remember fancy words to use to make great images, as generations you make has generation information with prompts baked in the image, like model information and whatnot in the image itself, so i load the image into notepad, and copy the information from there :P Image first mentions positive prompts, as in what you want it to do, negative prompts which is what you want it not to make/do, then seed (like in minecraft, every unique world has their own "barcode/serial key") to that specific result, models used and so on, and copy that to get it closer to what i want :P

Rest of the gibberish below the actual words is the regular image data itself

[half island]

Aye :P I tried to use a A6000 ada which generated stupidly fast in runpod, but could not for the life of me get it to use comfyui manager to fast install stuff -.-
So i need to learn how to do the linux stuff, like "pip install X && "cd into directory of extensions" git clone "link to extension" and same with fetching models, all in a single large command to automate the install. As it costs money to hold the data for you till next time

[meager carbon]

this is just the one that I have at home, not even counting the one i have in my office... but no I was joking, I didn't intend to really be serious when I said "weak"... I meant to say it as a joke.

this is raid1x3 ^

[cloud topaz]

hey guys is it normal that all my results faces look goofy? 😭
im using juggernaut v8 ith 25 steps and 8 cfg and 512x512

[karmic crown]

install a1111 and civitai browser helper, and it will download models for you and put them where you want them, then just rename extra_model_paths.yaml.example to extra_model_paths.yaml and set a1111 paths, and use models in comfy 🙂

no need to download models with curl or wget 🙂

[karmic crown]

if you are using default 1.5 model, it is very likely that you get something like that if your prompt is just a basic "girl in woods"

[cloud topaz]

oh really

i ll try a prompt from civitai

[meager carbon]

try a dif model

[cloud topaz]

also tried with realism engine

maybe my files are broken or smth?

[meager carbon]

doubtful

[half island]

A111 is quite slow compared to comfy, and comfy is also quite more effective where instead of having to redo base res image to then upscale, in comfy, i can set seed fixed for base image, then make it only redo upscale over and ovr until i'm satisfied. I do use auto for controlnet pose specific stuff though as i like it's way of handling it better

[karmic crown]

i said to use a1111 to download models with civitai browser (since that is much easier than to do that manually). than just set comfy to use models from a1111

[half island]

Remember to define the face. "pretty girl, detailed eyes, great smile, cute button nose" for instance in positive, as well as "bad face, mutated, floating limbs" etc in negative.

[cloud topaz]

okay thank you

it works better now! my bad guys

if any of you got any idea why the face is so smooth when face swapping i would like to hear it

[half island]

Oh true. Then i'd need said command do to that as well. where all is CLI command controlled until all is ready :P

[cloud topaz]

still the same problem of smooth face

[half island]

Add "detailed skin, realistic skin, skin pores" in positives for instance :) Alternatively, use a "realistic skin" lora

[cloud topaz]

you think it will have an impact on the result after face swapping?

currently i am generating a random picture with a girl and face swapping with a face

if i add lora and what u told me before the face swap it will have an impact?

[karmic crown]

and i have all models, loras... in comfy as well

Adding extra search path checkpoints /Users/viking/stable-diffusion-webui/models/Stable-diffusion
Adding extra search path configs /Users/viking/stable-diffusion-webui/models/Stable-diffusion
Adding extra search path vae /Users/viking/stable-diffusion-webui/models/VAE
Adding extra search path loras /Users/viking/stable-diffusion-webui/models/Lora
Adding extra search path loras /Users/viking/stable-diffusion-webui/models/LyCORIS
Adding extra search path upscale_models /Users/viking/stable-diffusion-webui/models/ESRGAN
Adding extra search path upscale_models /Users/viking/stable-diffusion-webui/models/RealESRGAN
Adding extra search path upscale_models /Users/viking/stable-diffusion-webui/models/SwinIR
Adding extra search path embeddings /Users/viking/stable-diffusion-webui/embeddings
Adding extra search path hypernetworks /Users/viking/stable-diffusion-webui/models/hypernetworks
Adding extra search path controlnet /Users/viking/stable-diffusion-webui/models/ControlNet

[half island]

I know where models goes :p I need an initial cli command that fetches all in one go, then I can edit auto's webui-user.bat with the custom locations afterwards. And once it's done, comfyui will then be up and running. As well as same for automatic to load into auto to fetch models, and comfy already got all the nodes that the initial command fetched

Got my local auto to fetch Lora's and regular models from comfy's location

[half island]

Though I will in fact need that as upscalers, vae's and whatnot is still copied over to auto as I haven't been arsed to check how I add that to the bat file lol

[feral mango]

is there someone very expert in automatic1111?
I get anyworkflow 100% identical with comfyUI
but AUTOMATIC1111 i never get the same iamge as the seed when compiying images and prompts
not even close
for example this model
https://civitai.com/models/77446/28d-stable-best-version
on the left the image reference, on the right my output

[half island]

Open OG image in Notepad and make sure you use the same VAE and sampling method as well

[feral mango]

absolutely all this is the same

[vocal burrow]

Why do I get different results from civit.ai examples :

1/ --xformers (and some other flags but mainly this one) will make stable-diffusion non deterministic. Meaning 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

should still apply to your situation (to some extent)

I wouldn't waste too much time trying to reproduce the exact same output.

[karmic crown]

also, Mac user will never get the same images as PC users with GPU

[vocal burrow]

different rng

different implementation for sure too

[feral mango]

1- its an issue since any iamge data does not have if --xformers were used
2- yes always i check it
3- yes send text2img keeps all the settings from official iamges that creator uplaoded for the model
4- no any overrides- added iamgen
5-- sure and same as point 1---
6-- not added to the oroginal text2img that i am testing i keep originall alwauy
7--not always the same seed as text2img
8-- yes this is something i have noticed that very easily you can maek a mistake but if you chose the right tab model on civitai it shows only those iamges for that model version

i am using citaibroweser + and you chose text2img directly from the img associated to that mkodel so there is no mistake

[karmic crown]

i am almost sure that there is no difference between gpu/cpu on mac, but i will try and let you know later

i have a headche now i need to go 😦

[feral mango]

this also seems to be a bug in current version of autoamtic1111
this really does not work and you ahve to choose up here

things are developing too fast and updates of code cant keep up

replication is only now 100% in comfyui

[vocal burrow]

1/ yup that's why I'm telling you to not waste too much time on it... Even if you knew if they used xformers.... they might have a different gpu than you => thus different xformers optimization => thus different results. As xformers act differently across multiple gpus.
5/ You'll never get every information and parameter from the metadata of an image. Sometimes you have enough.... sometimes not....
8 / I'm talking about auto1111 version, not the model.

Also I think ComfyUI wieght prompts differently. No idea how to translate ComfyUI weights to auto1111 weights. cf https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb's code for more information. Good luck

Sorry that's not the answer you want to hear.

[feral mango]

no its okey, i just dont want to as you said... not to lose time thinking there is something not set up properly. when there is nothing to set up

also, interestingly it seems at the end AI... is going to be unique at some point as well

[vocal burrow]

what do you mean ?

[feral mango]

since there are parameters imposible to save and to replicate... and even with that there are... lets say parameters that can not be parametrized?

[vocal burrow]

oh by unique you mean "hard to reproduce"

[feral mango]

if i dont get the iamge as the iamge reference that does not mean i am missing something to reproduce in the same style

yes xD

[vocal burrow]

Basic stable diffusion is deterministic

So you'll always get the same output for the same input.

But yes the more optimizations, modules, extensions, etc you add... The more likely it is someone in the chains didn't care about determinism

[feral mango]

can you replciate this one?
https://civitai.com/images/4198764

no lora no nothing

just the VAE

model and vAE

i was able to almsot replciate this one
https://civitai.com/images/4198753

feels like its missing something

[vocal burrow]

most likely incorrect VAE (due to coloring) + different xformers version (due to minor differences, (crowd in the background, etc)

[feral mango]

i am not using --xformers

and the VAE is the same

https://civitai.com/models/118561?modelVersionId=131658

[vocal burrow]

which VAE are you using ? screenshot your webpage

[vocal burrow]

the whole page

[vocal burrow]

this is using vae-ft-mse-840000-ema-pruned.ckpt

(also that's not the whole page)

[feral mango]

okey cant beleive i missed the VAE

the right one

now the colors are more close

is there a better website jsut for VAE?

[vocal burrow]

most vae (and models) are available on huggingface

[feral mango]

but like a collection of all them to download in bulk?

[vocal burrow]

not that I know of, and some could have different licenses.

there are not hundreds of them.

[feral mango]

ok thanks 😄

[vocal burrow]

that's what I have

much more discord TOS friendly

getting "closer without the quantization in K samplers"
I'll give it a shot without xformers for the lolz

[vocal burrow]

nope not happening, I cut my loss here, don't want to waste more time on this. Maybe they're using some embedding, maybe they have some ENSD value, maybe etc

[half island]

Set lora strength lower to minimize impact if it alters too much. Add lora between checkpoint and ksampler.

[umbral plover]

Any way or online service for upscaling images (or speeding up ultimateSd)? Last upscale took me 45 minutes but I'm still tuning parameters so if it takes that much time it's kinda hard (from 2k to 4k)

[half island]

Try upscayl, and how it's results looks.

[umbral plover]

It's definetly fast, but i don't see much difference before and after. Anyway, thanks, definetly will come in handy

[frigid notch]

I just downloaded the latest version of git and python for SD. But it ask me to update the pip install and upgrade, how do I do that?

[vast crest]

hi, guys, do you think I should use this setting?

Just out of curiosity, is it useful or not? 🙂

[ornate elk]

On auto1111 it won't do anything

[ornate elk]

Your python version is to new

SD only supports Python 3.10.6 up to 3.10.11

After uninstalling python and installing the other version you also need to delete the venv folder before launching

[potent bison]

does anyone know how the painter node works, i just dont get the interface and neither the image

[vast crest]

Ok, is there a way to update pyton and xformers via CMD?

[ornate elk]

Python needs a installer.
Check your xformers version first. At the bottom of the webui I'm browser

[vast crest]

python: 3.10.11  •  torch: 2.0.1+cu118  •  xformers: 0.0.20

[ornate elk]

Then your updated to the latest version

[night orchid]

hi ! I had to install stable diffusion webui again, now I get this:

I can't fix it

[vast crest]

ok thx 🙂

[night orchid]

I installed python 3.9

probably also because I have 3.12 still in the system ?

[umbral plover]

Tbh, I was tired of problems with python as i was trying diferentu ui's and i stumbled uponk https://pinokio.computer, don't know what reputiation it has but i can say that it takes aways process of installing everything and also for most ui's (except invokeai) it shares models so i don't have 3 copies of each model on disk

[night orchid]

@umbral plover thanks

[delicate berry]

Can somebody help me with python? My current problem is that whenever I try to run webui-user.bat it says it cant find python! plus: Whenever I try to check the currently installed pyhton version (python -v) with the windows command-line it also cant find python. Except I navigate within the windows command-line to the folder pyhton is installed in, only the it puts out the current version. ?

[night orchid]

have you set path ?

[umbral plover]

Sounds like its wasn't added to sys PATH

[delicate berry]

You mean the ticket box when installing? cause I always pick both??

[night orchid]

yeah it might have not done it, but check

[umbral plover]

Don't remember, you can add it manually https://datatofish.com/add-python-to-windows-path/

[night orchid]

there you are

[delicate berry]

Ok thanks a lot. I will try that later!👍

[ornate elk]

Your python is to new

Only 3.10.6 up to 3.10.11 is supported

[night orchid]

yeah I knew that, I installed 3.10 but it still looks for 3.12, so I had to rename the 3.12 folder to 3.10

its working now

thanks anyway

[umbral plover]

Yeah, if u have multiple versions it can be tricky

[cloud topaz]

okay thanks, any idea of each lora to use? i checked on civitai (or idk how its called) and the loras seemed pretty bad?

[half island]

Sort by popularity/downloads. And try one of them and see how the results will be

[cloud topaz]

okay! i am trying with ip adapter rn to see how i tlooks
thanks for your help guys i appreciate it

[half island]

Gotchu! I got 1600 lora's myself that i get a new whenever "i wanna test that" whenever i visit civit lol

[cloud topaz]

ahahah

[fiery dome]

Hi! This will be a very open question. I've been trying to make maps with added figures/objects/scenes by using img2img upload in Automatic 1111. I kind of like the scale of CFG 30 and denoising 0.5 but I would like more detail and definition....Any recommendations are super appreciated!

Settings:
Crowd of people eating inside their cars.
Steps: 50, Sampler: DPM++ 2M Karras, CFG scale: 5.0, Seed: 2758978573, Size: 1024x1024, Model hash: 31e35c80fc, Model: sd_xl_base_1.0, Denoising strength: 0.25, Mask blur: 4, Script: X/Y/Z plot, X Type: CFG Scale, X Values: "5, 10,15, 20,25, 30", Y Type: Denoising, Y Values: "0.25, 0.5, 0.75, 1", Version: v1.7.0

[astral comet]

Noob question, when i got a txt2img I really like, but some details are wrong (like hands, weird textures in small areas, etc), is there anything I can do in "post process" within stable diffusion to fix these flaws, or do I just have to learn to prompt better/use better embeddings? Is it through inpainting that you correct flaws?

[solemn moon]

correct, as well as control net, but i'd suggest inpaint and contrrol net on SD1.5 and not XL, using older models seems to respond better

or Photoshop

[astral comet]

yeah i'm already very comfortable with photoshop, digital painting as well actually. That said do you think it's still worth learning inpaint and control net?

[solemn moon]

defo's, there's like this feeling of difference when its generated and "never seen before" as apposed to copy pasta, but for each their own, maybe you have time constraints or the like..

also, why not add more skills to your skillset bruv

[astral comet]

controlnet seems very interresting, thanks

[noble anvil]

Anyone knows if there's an auto mute for either group or nodes in Comfyui? If you have different checkpoints it would be less aggressive on the GPU if you mute the checkpoint you're not working with if I understand correctly?

[woeful sedge]

hello, so im new to stable diffusion and got this blue yellow green on my image, the dark sushi is find, but other thing isnt, is there any solution for this?

[umbral plover]

One is 1.5 model, others are XL, check in descrption of XL models what settings they recommend, additionally, from what i read (not 100% sure) embeddings for 1.5 are not good for XL models

[ornate elk]

Hey, thats a VAE issue.
SDXL needs a different VAE

[woeful sedge]

okay got it

[woeful sedge]

so which vae should i use

[ornate elk]

This for example:
https://huggingface.co/stabilityai/sdxl-vae/tree/main

[woeful sedge]

and on the civit ai sdmodels there's a hash auto v1 that can be copied, what is it? is it important?

[woeful sedge]

many thanks, will try it

[ornate elk]

Auto v1?

You have an example of that ?

Oh you mean the Hash number?

[woeful sedge]

yeah hash number

[ornate elk]

That can be ignored

Its a unique ID for every model

[woeful sedge]

ahh okay, many thanks for all of the information 😊

[ornate elk]

No problem 🙂

[woeful sedge]

hello, so i do the png image information, to get the prompt from ai image that i just generated, but its not showing it, is there any solution for this

and for the upscaler, i want to install 4x-upscaler to change the latent, is there any easy way to install the 4x-upscaler, like the method when install/uploading stable diffusion models, lora, and vae

[ornate elk]

make sure you uploaded the photo from the outputs folder into the PNG-Info tab.

[ornate elk]

you can download upscalers (.pth files) and put them into the models/Esrgan folder. Thats all you need to do

[woeful sedge]

ahh so if i just screenshot some ai image, i cant got the prompt on the png-info

[woeful sedge]

okay got it

[potent bison]

how do I inpaint in comfyUI, i place the image and the mask and when i save the changes of the mask, i get an error

[karmic crown]

hush number can be used to verify if the downloaded file is the same as the one on server (you can check if they have the same hash), but as CS1o said, you can ignore it

[reef smelt]

Hires. fix makes my generation much slower, is there another method that is similar but faster?

from 3it/s to 3s/it

[ornate elk]

SD upscale script in img2img

[reef smelt]

Is it faster or is it an alternative?

[ornate elk]

its faster and an alternative

[reef smelt]

I'll check it out

[ornate elk]

uses the same technique as hires fix

[reef smelt]

is there a way to automate it?

[ornate elk]

you can batch upscale from a folder so yes

[reef smelt]

oh okay

nice

[reef smelt]

I tried it but it seems to generate like the same thing twice

[ornate elk]

you need to set the resolution to 512x512 (thats not the output resolution)
then set the denois below 0.3

then it will work

[reef smelt]

hmm oaky

[cloud topaz]

hey guys! I'm using ip adapter to face swap as i've heard it's consistent
have i done something wrong?

[proper pike]

What am I doing incorrectly that my model looks so bad? https://www.youtube.com/watch?v=pJ7uv32XtUw

[ornate elk]

Hey, you need to use sdxl IP-Adapter when using sdxl models

Also you need to crop the face more in the input image

[cloud topaz]

okay thanks for te advice i will do that 🙂

[wind sun]

so is there a way to faceswap videos without media encoder

[hot spindle]

Small question, are there diferences between Controlnet models wich are IP-Adapter and others that arent? Or Maybe I'm confusing what an Ip-Adapter is? I thought it was another word for Controlnet models

[ornate elk]

IP-Adapter is a controlnet model. you need to download it and then its usable in controlnet and SD:
these are all controlnet versions that need a model. Only Reference and Revision do not.

there are different model types for sdxl and 1.5 based models

[hot spindle]

ok, so I wasnt wrong...IP-Adapter is controlnet, but are all models considered IP-Adapters?

[ornate elk]

no

IP-Adapter is a type of controlnet

like openpose model or depth model

[hot spindle]

ok,never used that one then, its like prompting trough image right?

this one,right?For SDXL

[ornate elk]

yes thats the one for sdxl

[hot spindle]

ok,tks for the help, gonna investigate.

[ornate elk]

IP-Adapter is like recognizing whats in the input image and will try to reproduce that while generating the output image

there are different types

some work only for faces and the others for the whole image

the one from your screenshot will usethe whole image as reference

[hot spindle]

tks , i dont do people but saw a few other interesting aplications

[wind sun]

Hey, hello can i ask why when i try to generate a sport car or expensive one or as example ferrari its generating some red car but not a ferrari

[hot spindle]

This has very nice possibilities , instead of trying to prompt backgrounds

[wind sun]

haha i think that answers my question at some point xD

im starting to think comfyUI is better than this im using right now

[sonic ferry]

I'm having a hard time upscaling and it not looking out of focus or pixelated. Got the first KSampler set to 512x512, DPMPP_2M, Karras, 20 steps, 7cfg, denoise 1, going to a 'Upscale Latent By' set to 2x Bilinear hooked up to a second KSampler set to basically the same except denoise set to 0.56.

[rose skiff]

Can I make my pc do some kind of realtime img2img with my camera as input while I am on a zoom meeting?

I have a 4070 gpu, 14900kf cpu, and 64gb ddr5 ram

[cobalt shale]

Trying to train a lora in Kohya with a different base model than SD1.5, but when I choose "Custom", and open the directory, it shows no model for me to choose from. Could anyone please tell me how to use your custom model for Kohya lora training?

[sonic ferry]

Upscaling, right now I'm doing 512x512, running it through HighRes at 1.5x, and then the output to UltimateSDUpscaler set to 2x using 4xUltrasharp model, I should have set it to 1.35 or 1.5x instead to get closer to 1024x1024. But anyways, 4xUltrasharp, it being 4x means it takes longer? Are there others I should be trying? https://openmodeldb.info/ I'm looking through here but there are so many and I don't understand the 1x, 2x, 4x if I still have to set a size in UltimateSDUpscaler.

[fair oxide]

how do i get better results with amd gpu? if i copy generation data 1:1 from civitai i get way worser results with same model, loras and settings. is there any thing i can do about that?

[hot spindle]

@ornate elk Sry to bother you once more, do you happen to know the diferences between the two?In a nutshell...

Found this: https://www.reddit.com/r/StableDiffusion/comments/16gkmdx/confusion_with_different_sdxl_controlnet_models/

Sry, followed the link it perfectly explains the diferences: https://stable-diffusion-art.com/controlnet-sdxl/#Canny_models

[ornate elk]

thanks for sharing

couldnt have explained it like that 😄

but in fact these are just different variants from different research teams

[hot spindle]

I have read that page probably 3 or 4 times before..but just made sense now 🙂

[hot spindle]

Also,makes for for great read if you're into it: https://ip-adapter.github.io/

[heady sorrel]

Yes but why i have problems with SDE?

[sonic ferry]

How many models is it gonna load, I would think it's just 9 for the 9 tiles but I'm already at like 17....

Should I be cancelling this? I've only tried UltimateSDUpscaler once before and I didn't even have the right files then.

[hot spindle]

Its 9 tiles, check in console how fast is doing 1

But here's probably a mismatch there, you probably should have used tiles same size as image size

would be 4 tiles

[sonic ferry]

I'm in CPU mode, one tile is about 10-12 minutes.

[ornate elk]

What's your resolution?

[hot spindle]

ouch...its gona take a while then

[ornate elk]

Of the tiles

[sonic ferry]

I started at 512x512, but I have the Highres Script attached at 1.5x, so 768x768, Upscaler set to 2x so I was thinking it'd just be that 1536x1536, no?

Unless I'm misunderstanding, tiles set to 512x512, so shouldn't that just be 9 tiles it's gotta generate?

[ornate elk]

We would assume so. But there is also the "Tile Overlap"
Where the tiles overlap will be a smaller tile created to not create Seams

But also you should first try the ultime upscaler without controlnet tile.
Its not needed imo. But you can try and compare

[sonic ferry]

:/ I've just been playing with different upscaling ways, Highres fix alone, Upscale Latent By, and now this...which I didn't think would take this long when I saw 9 tiles. 😛

[ornate elk]

Without controlnet it would be faster for sure

[sonic ferry]

Ah, I'll try that next, I was just following what a guide said.

[hot spindle]

@sonic ferry but dont you have GPU of any sort?

For a simple 2x upscale might be faster to use ChaiNNer

[hot spindle]

Didnt understand that, thought you were starting with 768x768 (SD 2.0 or something like that) in that case yes, probably better to stick with the same tile size for the model

Its funny because I'm experimenting with 8xNMKD_Superscale_15000_G , in ChaiNNer it strictly follows model, so no option to do just 4x , but with ultimate SD Upscale its limited to 4x , so it does 4x then 2x from that 😛

[umbral plover]

Anyway to use only one node of image save to save all connected images? (WAS Suite Image Save node) // edit found Batch Image

[sonic ferry]

No GPU, it's just one of these. I'd like to try the 750Ti SC if I had a way to hook it up just to try, that's in an old desktop that is very slow now. LOL

[hot spindle]

Not sure about A1111 as there was a crackdown from Google, but have you looked into runing ComfyUI as backend in Colab?

[grizzled merlin]

Does anyone know how to solve the gpu spike when running SD?

[worldly rain]

Hey guys I have a problem, I generate the image I want with hires fix but it's almost getting crashed everytime. So I wanna generate the image without the hires fix first, then I wanna send it to img2img to upscale it but I don't see the hiresfix in img2img window. I can use other upscaler yeah right but I have some particular settings I wanna apply with hires fix, so is there any way I can use hiresfix on img2img

[marsh arrow]

It says I have a pending request in bot 6 from hours ago. Should I tag a mod?

[woeful sedge]

hello, why my open pose editor pop up is so small

[covert fox]

how to fix this?
on google colab

[ornate elk]

You can enlarge it while holding the corner on the down right

[ornate elk]

You need to update the webui files

With git pull

[ornate elk]

Make sure to set Hires Steps to 10 and upscale not higher than by 2, so it won't crash

[worldly rain]

yeah the mistake i did was generating the image in 1024x1024 then upscaling it 2x, i did 512x512 and upscaled 2x and it worked withouth crashing

I use sd 1.5

thanks 🙏

[woeful sedge]

cant do that, any other solution? and how to change my ui to english language, because i change this thing, but it still on chinese

[karmic crown]

maybe problem exists in all samplers, since that sentence is very confusing
i have never noticed that problem since i never used more than 30+15 steps
just lower the number of steps

[ornate elk]

@worldly rain
No problem 🙂
If you want to use SD upscale script in img2img then you need to set the denois below 0.3 and the resolution to 512x512 and then choose an upscaler

[karmic crown]

some samplers "converge" (reach a stable state), while others do not
https://diffute.com/glossary#converging

with samplers that "converge" you will get almost identical images after a certain number of steps, so adding more steps is usually just a waste of time and resources
with some samplers, you can even get a worse image if your step count is too big

here you can see how different samplers behave with different number of steps
https://www.artstation.com/blogs/kaddoura/pBPo/stable-diffusion-samplers

in most cases, "converge" usually occurs around 40 steps
so I would never understand why someone would use more than 60 (40+20) steps, probably because they are not aware of this

[ornate elk]

What webui did you installed?

I know there is some Chinese fork of auto1111

[karmic crown]

might be a bug in lobe theme, try to update

[woeful sedge]

sd_webui_aki_v.4.6, it has a chinese title to run the sd

[karmic crown]

in that case your problem is a1111 fork

[woeful sedge]

i think i already updated it on yesterday

[ornate elk]

Ah okay, idk if the fork supports a full English ui

Have you tried to restart the ui after installing the language?

[woeful sedge]

im not installing any language yet.... maybe thats the problem?
i just change the theme settings to english

[ornate elk]

Ah okay, then you need to click on extensions tab. There click on Available. Then click on "Load From"
Then tick the Language box and install the English pack

Then restart

[karmic crown]

fork probably has a setting that overrides the one selected in theme

[karmic crown]

or this

[woeful sedge]

which one is the language box?

[ornate elk]

Localization

Untick that

[woeful sedge]

okay

[ornate elk]

Then search for english

[woeful sedge]

already entered but nothing show up

[ornate elk]

Thats strange, maybe only type EN

[woeful sedge]

same

[ornate elk]

Gonna start my webui and check

[woeful sedge]

okay thanks for helping me 😄

[karmic crown]

go to settings

show me your main menu

[woeful sedge]

this one?

[karmic crown]

here

[ornate elk]

Np but it seems like there is no English language pack as the original auto1111 is already english

[karmic crown]

exactly

put it to none

that is chinese

[ornate elk]

Change that to English or None if possible

[karmic crown]

none is enough

[woeful sedge]

reloading

nothing change

[karmic crown]

i just checked, there is no english

[woeful sedge]

is there an easy way, to install english language, and change those to english

[karmic crown]

but

[woeful sedge]

copy .json to some folder maybe

[karmic crown]

go here

disabel this one

[woeful sedge]

okay brb

[woeful sedge]

reloading

[karmic crown]

and? 🙂

[woeful sedge]

same 😄

[karmic crown]

green one

[karmic crown]

what do you have there?

[woeful sedge]

change it to zh-hans?

[karmic crown]

no

[woeful sedge]

none

[karmic crown]

zh-hans is simplified Chinese

leve it on none

[woeful sedge]

yeah its on none

and no zh-hans option

[karmic crown]

so, you disabled extension and this does not show 4 versions of chinese translations, and your ui is still in chinese?!

this is very strange

[woeful sedge]

i think i will try this

[karmic crown]

actually do the oposite

go here

and remove everything except this

and restart

does it work?

[woeful sedge]

this not work, nothing more options on language setting

[woeful sedge]

okay will try this

just have 1 file on the folder

[karmic crown]

i have no more ideas then

except to install a1111 manually

[woeful sedge]

if i install another a1111 will the plugin applied to the new a1111?

[ornate elk]

which plugin?

[woeful sedge]

all of the extraction, lora, sd models, embedding, etc

[ornate elk]

you can copy/move them over

the extensions need to be reinstalled in the other version

the Theme extension is called Lobe Theme

[woeful sedge]

okay got it, thanks for the info

[ornate elk]

no problem, for a detailed install guide you can checkout the Pinned Messages of this Channel

[unique anchor]

nvm figured it out

[potent bison]

I instaled manager in ComfyUI and intstalled a custom node and when try to restart doesnt open anymore, and I deleted the custom nodes and this error appears

if i put the custom nodes i get a larger error

[ornate elk]

would not recommend installing comfyui with stability matrix

easier is just following the guide on the github from comfyui

[umbral compass]

I'm using ComfyUi and question regarding ControlNet if I also have Auto1111.

If I put my Model or for example "depth model" in the "stable-diffusion-webui/models/ControlNet" folder it shows up for ControlNet in ComfyUi.

But if I put it in the "stable-diffusion-webui/extensions/sd-webui-controlnet/models" for Auto1111 it doesn't up for ComfyUi.

So does the folder really matter on where I put my Model if I use both Auto1111 and ComfyUi?

[astral comet]

So i've tried creating a model with pictures of myself, following two different tutorials that had me go through a process within a Google Colab document. In both instances, I got blocked somewhere along the way because of some thing being outdated, or incompatible, or whatnot. Both tutorials were over a year old, anyone knows a more up to date tutorial?

[fair oxide]

how do i get better results with amd gpu? if i copy generation data 1:1 from civitai i get way worser results with same model, loras and settings. is there any thing i can do about that?

[vocal burrow]

It depends what you mean by better. Do you meant faster ? Or closer to what's displayed on the civitai picture ? (Or both?)

[karmic crown]

since he wrote "worser results with same model" i am almost sure he was talking about the quality

[vocal burrow]

probably just want to be sure before I copy paste a wall of text :p, don't want to scare them

But now that I've spoiled the surprise

Why do I get different results from civit.ai examples :

1/ --xformers (and some other flags but mainly this one) will make stable-diffusion non deterministic. Meaning 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

[fair oxide]

technically if i paste the exact generation data with the same seed i will get the exact same result right?

[fair oxide]

since i have an amd gpu i cant use xformers

i have installed a fresh installation without any settings i still get the results

i can show you some results let me check

[mighty sorrel]

i am having problems in installation of stable diffusion on amd gpu MAC

or someone can share a step by step tutorial.

[fair oxide]

.

[fair oxide]

i get results like this

this is not from civitai but its a good example

this is not that bad but not the result i expected

[vocal burrow]

Well like I stated in the daunting wall of text, if you're only copy pasting what you see on the civitai webpage then you're missing a bunch of meaningful parameters.
(Also yes, you can't use xformers but that doesn't mean the person who posted the image didn't use it)

[fair oxide]

so there is nothing i can do

its not only about not having the same results as posters on civitai, my results are just bad

[vocal burrow]

show an example of what you're trying to reproduce (civitai page) and the results you get (the actual output .png and a screenshot of your txt2img page showing every settings)

[fair oxide]

most of the images there use hires fix wich i cant use so its not the exact result but i will check real quick

and civitai takes like 5 minutes to load

bruh

[vocal burrow]

yeaaaah civitai is not the most stable website out there.

it works (slowly) on my end tho.

[fair oxide]

thank god someone finally told me its not my fault

i thought i blocked some ports on my router 💀

[tulip saffron]

Hello, I would like to implement something like what Leonardo has in their Image to Image in SDXL with Automatic1111. I imagine that this involves ControlNet? How may I use a float value to control how much the AI is affecting the Image Prompt? Thank you!

[old niche]

你

[vocal burrow]

from civitai's discord

[vocal burrow]

私

[fair oxide]

i just switched from --medvram to --lowvram now my results are better, but it takes longer (like 30 minutes for a 512x512 20 step generation)

[vocal burrow]

strange, it shouldn't impact the output of the image.

[fair oxide]

i generated the same image i can send it

they do look different right

or i am just delusional

[fair oxide]

wait i just made this image and it looks kind of good

and it was possible to generate 520 by 840

[vocal burrow]

they look exactly the same

[fair oxide]

okay your right

[tulip saffron]

Using SDXL with Img2Img and I get this error: 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.
Time taken: 3.2 sec.
I have a 4090 laptop GPU...does this need the --no-half argument?

[vocal burrow]

like you've said it probably involves controlnet and/or the CFG scale, without the source code you can't be sure what's the special sauce underneath.

[vocal burrow]

what's your gpu and command line args ?

[tulip saffron]

Right. But do you or anyone else have experience with setting up an Img2Img with a "Strength" variable? I'm aware of Automatic1111's Denoising Strength, but it seems like it doesn't work as well without ControlNet and control over colors.

[tulip saffron]

Thank you for the reply. As mentioned in the original message: 4090 laptop GPU and set COMMANDLINE_ARGS=--autolaunch --update-check --xformers

[fair oxide]

i get this when i use tiled vae (the black thing in hear face isnt wanted lmao)

[vocal burrow]

""""strength"""" can be cfg scale

[tulip saffron]

I'm now trying set COMMANDLINE_ARGS=--autolaunch --update-check --xformers --no-half-vae which seems to be working

This Img2Img in SDXL/Automatic1111 doesn't look good right now. I'll have a play with Denoising Strength and CFG Scale, but the first image is terrible

[fair oxide]

why does it always do this?

[vocal burrow]

what are your command line args ?

[fair oxide]

--use-directml --medvram --opt-sub-quad-attention --opt-split-attention-v1 --no-half-vae --upcast-sampling --lowvram --no-half

[tulip saffron]

CFG Scale 10 👎

[vocal burrow]

cfg scale 10 is way too much

[tulip saffron]

Yes. It was 7 in the original, which also looked bad

CFG Scale 3.5

[vocal burrow]

what are the settings you're using ? screenshot your webpage

[tulip saffron]

For reference, this is what Leonardo's produced, which I love

[vocal burrow]

ok, good luck trying to reproduce it

[vocal burrow]

try with a encoder tile size set at 256

[fair oxide]

okay

[mighty sorrel]

torch not compiled with cuda enabled mac stable-diffusion

any solution?

[fair oxide]

what is you gpu?

[mighty sorrel]

amd

[karmic crown]

wait

[karmic crown]

mac?

[mighty sorrel]

yes

[karmic crown]

in that case its the same for any gpu

[mighty sorrel]

kindly give me any solution.

[karmic crown]

and mac does not support cuda

we do not have that

this is perfectly normal for Macs

no module 'xformers'. Processing without...
no module 'xformers'. Processing without...
No module 'xformers'. Proceeding without it.
Warning: caught exception 'Torch not compiled with CUDA enabled', memory monitor disabled

[fair oxide]

still the same

[mighty sorrel]

any solution to use gpu for stable diffusion. generation is very much slow

[karmic crown]

it can be slow even with GPU, what is your exact mac model?

[vocal burrow]

and without tiled vae ?

[karmic crown]

i have 2 macs
iMac Intel 16 GB + Radeon 8GB
M1 MacBook Pro 16GB

[fair oxide]

without tiled vae i cant generate 1024x1024 but if i generate 512x512 it looks the same

its the same settings as the images i already sent

[vocal burrow]

by the same do you mean "with the grey square" ?

[fair oxide]

these

[fair oxide]

without it

[vocal burrow]

that's probably because you didn't use correct setting then

[fair oxide]

i used the same just higher resolution and tiled vae

[mighty sorrel]

Apple m1

[vocal burrow]

and what are those settings ?

[mighty sorrel]

macbook air 8gb

[karmic crown]

i was afraid you would say that

show me this part of UI
version: v1.7.0  •  python: 3.10.13  •  torch: 2.1.2  •  xformers: N/A  •  gradio: 3.41.2  •  checkpoint: 8b2c696413

[vocal burrow]

those are the tiled VAE settings... I want to see the settings you've used to get two similar pictures in img2img without tiled VAE

[fair oxide]

i didnt use img2img

[karmic crown]

and show me your webui-user.sh

[fair oxide]

i just used the same generation data as with the other images just turned on tiled vae

[mighty sorrel]

API  •  Github  •  Gradio  •  Startup profile  •  Reload UI
version: v1.7.0  •  python: 3.10.13  •  torch: 2.1.2  •  xformers: N/A  •  gradio: 3.41.2  •  checkpoint: 6ce0161689

[vocal burrow]

ok then try with encoder tile size at 512 and decoder at 128

[karmic crown]

this is fine, show me your webui-user.sh

[fair oxide]

ima try

[mighty sorrel]

ok let me show you.

[karmic crown]

those settings (especially tile size) should be locked with the message "i know what i am doing, let me see the settings"

[mighty sorrel]

can i send the whole file or i can paste the code here?

[karmic crown]

just drag & drop here

[mighty sorrel]

ok

[fair oxide]

before finishing its normal

[vocal burrow]

that's expected

[karmic crown]

just do this

[fair oxide]

bruh

[vocal burrow]

try using those command line args : --lowvram --no-half --opt-sub-quad-attention --opt-split-attention-v1

[fair oxide]

only these?

[vocal burrow]

yes

[fair oxide]

venv "D:\stable diffusion\stable-diffusion-webui-directml\venv\Scripts\Python.exe"
fatal: No names found, cannot describe anything.
Python 3.10.9 (tags/v3.10.9:1dd9be6, Dec 6 2022, 20:01:21) [MSC v.1934 64 bit (AMD64)]
Version: 1.7.0
Commit hash: d500e58a65d99bfaa9c7bb0da6c3eb5704fadf25
Traceback (most recent call last):
File "D:\stable diffusion\stable-diffusion-webui-directml\launch.py", line 48, in <module>
main()
File "D:\stable diffusion\stable-diffusion-webui-directml\launch.py", line 39, in main
prepare_environment()
File "D:\stable diffusion\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
Drücken Sie eine beliebige Taste . . .

[vocal burrow]

oh my bad yeah, add --use-directml

(they added it somewhat recently that's why I keep forgetting it.... not sure who would get the directml fork and decide to NOT use directml but whatever)

[mighty sorrel]

@karmic crown please check

[upbeat oracle]

hey community, i'm new here and have a little question (don't know if this is the right place thought):

If you have the choice between a GeForce RTX 3060 12.0 GB and a GeForce RTX 4060 8Gb for the same price, which would you choose (mainly for stable diffusion)? 😊

[vocal burrow]

There's no "right place" for this, it might be better to ask in #💬｜general-chat . It's a tech question but not a support one.
Anyways, it depends what you're gonna use this gpu for. generation ? training ? sd15 ? SDXL ? gaming ?

[mighty sorrel]

@karmic crown please suggest after checking it

[karmic crown]

find this line

#export COMMANDLINE_ARGS=""

and add this below

export COMMANDLINE_ARGS="--skip-torch-cuda-test --upcast-sampling --no-half-vae --medvram --use-cpu interrogate --disable-model-loading-ram-optimization"

run ./webui.sh

it should be faster, but do not expect is fast as some nvidia cards

[mighty sorrel]

@karmic crown thank you very much

[marsh arrow]

Still getting an error message when I try to use the bots.

[karmic crown]

is it better?

[vocal burrow]

so is it working now ?

[upbeat oracle]

i'm a bloody beginner, all of the above (want to test several things but the target later is to train). Currently i calc on my 8-core CPU notebook (takes ~1 hour for a pic) 🐒