getting ready to train embeddings | stable diffusion | automatic1111
Vložit
- čas přidán 31. 05. 2024
- a bunch of things to help in Stable Diffusion. use this video as a reference for getting started in training your own embeddings.
💲 My patreon:
patreon.com/RobertJene
🍵 Buy me a coffee:
www.buymeacoffee.com/robertjene
This is the part 2 video, how to train embeddings:
• how to train any face ...
If you are having trouble with the script that inspects embeddings:
• Analyze Trained Embedd...
00:00 - intro
00:39 - install and work with it first
01:08 - check your VRAM
01:56 - setup your batch files
05:05 - get models and embeddings
06:34 - upscalers
08:08 - VAE's
08:49 - change some stable diffusion settings
12:45 - create custom training template file
13:35 - install apps and repositories
previous videos:
Install Stable Diffusion on PC | Automatic1111 | 2023
• How to Install Stable ...
How to use Stable Diffusion on PC | Automatic1111 | 2023
• How to use Stable Diff...
webui-user.bat:
set COMMANDLINE_ARGS=--xformers --no-half-vae --precision full --no-half
webui-vanilla.bat:
set COMMANDLINE_ARGS=
Models:
huggingface.co/runwayml/stabl...
civitai.com/models/4201/reali...
Upscalers:
8x_NMKD-Superscale_15000_G
huggingface.co/uwg/upscaler/t...
8x_NMKD-Faces_160000_G
nmkd.de/shared/?dir=ESRGAN/Mo...
The popular VAE’s:
Anything-v4.0.vae.pt
huggingface.co/andite/anythin...
Grapefruit.vae.pt
huggingface.co/iZELX1/Grapefr...
vae-ft-mse-840000-ema.pruned.safetensors
huggingface.co/stabilityai/sd...
Settings stuff:
Settings - User Interface - QuickSettings List
sd_model_checkpoint, CLIP_stop_at_last_layers, sd_vae
images filename pattern:
[model_name]_[seed]
check the box:
Add number to filename when saving
Apps to download:
Irfanview
www.irfanview.com/main_downlo...
Gimp
www.gimp.org/downloads/
GPU-Z
www.techpowerup.com/download/...
Winrar
www.win-rar.com/download.html...
Github repositories:
My github repository (optional)
github.com/robertJene/stableD...
Uses a python file from here to inspect training:
github.com/Zyin055/Inspect-Em...
_____________________________________________________________
😆💗 Social Links:
Twitter: PerRevolutions
Facebook: PerRevolutions
Instagram: PerRevolutions
TikTok: tiktok.com/@perrevolutions - Věda a technologie
If you are having trouble with the script that inspects embeddings:
czcams.com/video/_AN6PpGNomw/video.html
your video is better than the last 50 i watched!
Come back later for the actual embedding video, it's being edited
dude this is a quality af tutorial
Thanks!
Thanks for the dollary doos
Maybe you'll come over to my new patreon, if you are interested
Patreon.com/RobertJene
awesome video man,
Love you bro, well done.
Great video buddy! Subscribed, look forward to your LORA vid.
hey rob, you will like this one, go into file explorer to a random path, then in the path textbox clear it and type cmd. You will open command prompt directly from that folder ;-)
No, and tbh I'm getting tired of explaining why you shouldn't do that
@@RobertJene huh? its the same as opening cmd and typing cd /d %path%. please eleobrate, and sorry if you've already mentioned it in a video. i havent seen you speak about that? I always thought it was one of those cool super hidden windows features??
@@traida111 I actually go over it in the video I am editing right now.
But it "breaks" the history of file explorer.
yes. File explorer has history like a web browser.
I think my head exploded
🤯
dude. LOL. check out my videos on LoRA's. they're 1,000% worse!
Thank you for this, well done. couple of questions, why leave "Use cross attention optimizations while training" on when some suggest it messes with training from the last update? and why clip skip at 2 instead of 1? also, isn't' xformers auto turned off on training?, which leads to my last q which is just why the change to the webui-user.bat lines in the vid from the description? Hope you have time for a brief response but either way this vid is really far superior to what's out there. I have run dozens of trainings by feeling around the limited info available and still, as you can see, I have basic questions. If I had watched this first I would have been so far ahead on the learning curve.
Use cross attention optimizations while training saves VRAM during training
xformers saves memory when generating images, on some models, but it crashes training embeddings
clip skip.... more explanation is in my tips video. but 2 is the most common best result when testing with generating images
about changing webui.user.bat, that is explained in the video.
you need ZERO added settings / arguments during training so that it doesn't crash or make weird results
FastStone Image Viewer better in terms of mass renaming, resizing, cropping and many more. It have portable version as well. I use it for years. Try yourself. I like hotkey X when cropping, and cropping is very customizable there. Also 'Copy to' and 'Move to' is useful.
Cool I used to use FastStone for making screenshots for technical documents
But is the image viewer free?
@@RobertJeneYes, FastStone Image Viewer is free.
Small Doubt.... when i run GPU_stats... my gpu memory which is 8gb shows half of it being used, around 3900 even though i have no programs running. could u please help me out as urs was showing only 415 mb used 17:40
yeah what you want to do is research how to use task manager to see what's going on.
Here's some tips to get you started:
1. press Ctrl+Shift+Escape
2. Find an option that that shows "more details"
3. Go to the details tab
4. Sort the list by what is using the most memory
It's not always bad, but sometimes it is, sometimes you need to uninstall something.
Other times you just need to remember to close certain things before running GPU intensive tasks
hello,
If i don't have Xformers, do i need to edit the notepad as "--xfomers --no-half" ? as argument ?
i run on 3090
yes.
I recommend using a separate batch file, webui-user.bat
quick question, does this not work on amd gpu? specifically rx6600 xt?
good question. at this time, image generation on AMD gpu's is not really working.
Look for my video titled "A better A.I. GPU for the money"
and I explain it there
your GitHub repo is nonexistent...can we get a link please?
That, and another GitHub linking my description got mangled by CZcams for some reason.
I just fixed them.
Hi nfixer 👋
nfixher?
@@RobertJene one of your embeddings
dont get it..., got better results without everything of that... just get proper models from civit ai, but also with base sdxl model...
just looked how to create a model from around 5000 pics of a person i got... :(
5000 pics of one person???!!!! 😳😲
@@RobertJene Yes Sir, it's a chinese famous professional Cosplayer, downloaded all freely available pictures and sorted out doubles...
@@hogstoothairsoft1967 dang...
Hey there, Sir
First off, I wanted to say how much I like your SD tuts. I learnt so much thx to them.
Now, you keep saying that people need an NV GPU... well for those with AMD GPUs here's a tut how to instal SD on AMD cards:
czcams.com/video/8xR0vms0e0U/video.html by The Kingdom
...worked for me and my SD runs WAAAAYyyyyy faster with this set-up.
Ofc, I do not try to promote anyone here or anything, just remove my comment if you feel like it, Sir.
Ken
I tried this version, and I had TONS of problems with it.
the main problem was that it uses 100% of the AMD GPU's VRAM, no matter how much VRAM is on the card.
I have a few from when I used to do crypto mining.
That plus stuff like hires.fix don't work
random crashing
etcetera
@@RobertJene so true, unfortunately. Very often I get the 'too little memory' (cannot allocate xxx memory). Yeah. But anyway it's slow progress anyway with 8gb vram only 😔
@@kenbismarck4999 Yeah I tested it on a 8gb card first, with the intention to make a video + help people with older AMD cards.
Then I tested it with a RX 6800 which has 16 GB VRAM and it was only slightly better.
yo dude ..this is the error im getting with you loss tool, its looking for a .cvs not vbs? ERROR: The file, "textual_inversion_loss.csv" was not found in provided folder, or it is not available:
H:\AiArt\embeding-training-data
That either means you gave it the wrong folder, option f.
Or the training hasn't outputted any embedding files yet, so the log hasn't started
@@RobertJene You nailed it. So i guess you start the tool after you start the embed?
@@hairy7653 basically you don't really start the tool, you just run it to check what the log folder has in it at that time when you run it.
So you don't sit around and just stare at it all day while it runs, you just check it every 15 minutes or so.
Everytime you run an option, it analyzes what has been done so far up to that point
@@RobertJene yo i gpt it working, sweet. But I'm getting errors on option 3. It asked for panda and now 'matplotlib'..i guess i need these too? sorry to be a pain!
@@hairy7653 weird. Maybe it's something you have to install with pip. Copy paste the whole error line into google
ahm. Tipp: you can write "cmd" in any windows file explorer at the address line on the top. This will bring up your command window in the chosen directory.
Lol I've gone over this so many times. That breaks file explorers history and I can't use back or up buttons after doing it