Textual Inversion Tutorial - Embeddings and Hypernetwork basics and walkthrough
ฝัง
- เผยแพร่เมื่อ 19 มิ.ย. 2024
- Hey Everyone! This has been a popular request in both comments and in the discord, so I put together a more comprehensive breakdown while focusing on both "how it works" and "how we make it work".
Textual Inversion allows us to push our models toward a very specific look, and here's where we can learn how to give that push.
00:00 Intro
00:59 Understanding Textual Inversion
06:10 Preparing Images
08:00 Creating The Tokens
09:22 Preprocessing Images
11:20 Training A Hypernetwork
16:37 Training Embeddings
17:05 Understanding Templates
19:05 Examples with a larger set
linktr.ee/FrankTankGaming - ภาพยนตร์และแอนิเมชัน
I wish it didn't take so long to find this video. Confirmed much of what I've learned the last two weeks and showed me clarity on a few things that I found confusing. Well done.
Sorry I couldn't help any sooner, but thank you for your feedback and I appreciate your support! :)
This is such an excellent video. I love that you have shared example outputs, settings and most of all, descriptions and explanations of underlying concepts & theory. Thanks
Thank you so much!
Thanks for making this great tutorial, everyone else wants to lecture us for an hour, you nailed it in 20 minutes.
Glad you enjoyed it!
8:27 The initialization text is used to limit which part of the image the inversion is applied to. * means all of the image will be influenced. For example if you set it to tree it will only draw trees in that particular style (or for subject it would add a tree where there was none).
For subjects keep it as basic as possible i.e. 1 word to limit the changes to other things in the image.
For styles you may want to keep the asterisk, or you can change this if there are only specific things which you want to have the new style.
This is the only video I found with information regarding the Hypernetworks training
This is the most helpful guide I have found! Thank you for taking the time
I noticed my train tab has Create Embedding, Create Hypernetwork, and Train. I don't have Preprocess Images. That whole section is gone
now the options of preprocess images are in "Extras". look there and tell me if its works
Same here, did you figure it out?
real good vid! very helpful thx :)
Thank you!
I saw people trained textual inversion models for negative prompts so people could be able to type in one thing like "easynegative" and it gets rid of a lot of problems that typically tend to come with generations. I am having some trouble with something and would like to train textual inversion very specifically for some negative things to see if I can get rid of these problems. I am watching this video trying to get it figured out, and I'm hoping it can work... if it doesn't, I'll be back.
Honestly that's the first I'd heard of that use case - I'm really curious what kind of training material would be used for something like that. If you haven't seen it, my video breaking down Loab and negative prompts might help you get a better understanding of how negative prompts work, as far as what I found with testing: th-cam.com/video/RYiXwKTVR2A/w-d-xo.html
@@FrankTheTankGaming I'll check it out for sure. I found out by trying to use someone else's lora as a negative prompt set that apparently loras don't work for it, so that saddens me :') I'll go look at it right away. Tyvm.
Nice! thank you very clear... Do you have some experience with LORA as well?? :) people with a clear voice and some deeper knowledge are important :) tnx for this!!
I haven't yet figured out a clear workflow for LORA, but I absolutely want to try to make a video about it once I have a better understanding of exactly what to do. Especially if this video does well, I want to continue to make more tutorials so I appreciate the feedback. 😃
@@FrankTheTankGaming Tnx! I like the way you teach... so thanks!
@@FrankTheTankGaming Simply DL Kohya GUI by Bernard Maltais, training workflow is pretty similar to this video's workflow for TI training. Those trained LoRAs work without any extra work in A1111 webui.
Goddamn. I am really glad you took the time to sit down and make that video. I saw a lot of things and they just didn't work. Something was off. Went back, learned.. trained the nets. Makes a lot more sense now.
Thanks, and I'm glad to hear that. A lot of these methods stop working relatively quick because things keep getting updated lol.
My SD (a1111 and Forge) doesn't have the preprocessing tab. :-/
Excellent video, informative and to-the-point. I am curious though, I notice that you're able to view the contents of text files without opening them. Unless you're on Linux I can't figure out how. Share?
I'm using a program called One Commander instead of Windows explorer 🙂 Hope that helps!
thanks for this tutorial. quick side question I couldnt find the answer to, how have you scaled down the previews of embedds, and hypernetworks in the text2img tab?
in the settings, find 'Default view for Extra Networks' and you have an option there for thumbs vs cards ;)
great video, with lots of helpful information and like others mentioned pretty unique in the flood of videos on that subject. Unfortunately on mac-os it's a bumpy road. I got stuck with "AttributeError: 'NoneType' object has no attribute 'get'" in terminal while training hyper network and the embedding will not show up in the dropdown menu, though it's in the folder... so that's the terminal station for me :-/
13:49 It's really cool to see, that - while close - this image generated is not just a copy! The angle is a bit differnt and it added wrinkles above the raised eyebrow. It's a pretty bad-ass result!
Maybe I should try hypernetwork, I had no success trying to train an embedding, even though I have a great graphics card and a new computer, stable diffusion Automatic1111 and I let it run for hours. The renders just didn't start approaching the target :(
I had 27 images at 512x512 and I used the stable-diffusion 1.5 pruned model.
I tried it several times, carefully followed tutorials - I have no idea why failure.
I'm new to this, but i suppose...when in doubt? Add more photos?
Let me know what you've discovered. =)
For some reason there's no preprocess tab anymore.
14:41 - what did you do to make your extra networks picker look like a popup window? Or is that by the magic of video editing?
Ah that was just video editing - there's only the two views that I'm aware of, where you can get a larger display card or a smaller one in the options
@@FrankTheTankGaming OK! I was thinking what the hell. Anyway, that window is horrible, would rather get such a nice looking popup! 😃
hi Frank, is embeddings tab bugged in auto1111 and Forge?
15:55 if you had the hypernetwork file in the correct folder, someone else's prompt image would work for you
Hello and thank you for these great videos, I was wondering if you had a tutorial which explains basics about how to download webui automatic 1111 as well as dreambooth, lora etc.. I also have no idea how python really works and how I know there is a correct path between Python and webui automatic, I have downloaded all of these from one tutorial but it feels messy cause I don't have deeper knowledge about these things.. I also have no idea how to make updates and how all of these things interact with each other 😢 I just dont want to simply repeat what people do without understanding it in a deeper level.. If you know a good tutorial or have done one, it would be wonderful.. thank you so much
You know what- I was thinking about creating an updated tutorial since my other one is getting a little dated... I think your comment just put me over the top for that.
Stay tuned because I do want to do another proper install tutorial with some of the new information and tools, and I'll do my best to make sense out of all the complicated stuff
@@FrankTheTankGaming That's fantastic ! Thank you so much :) its been a month and even tho I'm getting better, its still so fascinating and alien to me... I'm trying to run everything via Runpod like most people who do not have a powerful computer, my goal is to make an animation like in Corridor Crew for my music video clip. I have not seen any tutorial on how to turn a "video to anime" via runpod yet. So if you ever do it too, I'll watch it many times ^^
Followed steps to a tee but when I use the in my prompt, a completly different image pops out. I dont know if it matters but my training stops every 500 steps and says complete.
Also - it says training complete but cant find the sample images in the log file?
Bro you have cool voice
Great tutorial, I have started getting error
KeyError: 'clip_skip'
Anyone else get this?
So it's not possible to train an embedding that copies a face perfectly?
An embedding is probably the weakest form of training, so it's far less likely with an embedding than with a hypernetwork, for example. Dreambooth and LORAs give the most spot-on results. (Video tutorial coming soon on that 😁)
@@FrankTheTankGaming thanks for the reply! I've been trying to train an embedding of my face for days and it always looks like a caricature no matter what I do. I wish I had known it was this hard to make it copy a face right before I started 😩
@@babydollface Are you working exclusively with hypernetworks and embeddings? Did you try LoRa or DB?
@@joeyc666 I gave up on embeddings. I trained a lora of my face and it looks pretty good
@FrankTheTankGaming Hi ! I'm a newbie, so sorry if my question will sound stupid. I just came to your video after I watched this video: th-cam.com/video/dVjMiJsuR5o/w-d-xo.html
Here guy explains about 4 main training methods: dreambooth, lora, textual inversion, hypernetworks. Also he mentions about "aesthetic embedding" which according to him is totally useless. And I somehow liked "textual inversion" most, as it's small file and you can get high quality results. But I saw here in the comments, that you say, that "textual embedding" is the weakest training form. So now I'm confused: is textual inversion is the same as textual embedding ? Do they have anything in common with "aesthetic embedding" ? Which is the best method to train ? Before your video I thought about "textual inversion", but now I'm confused and linked more to Lora kind of...
'Textual inversion' is the official name of the process that can create both embeddings and hypernetworks. Hopefully that clears up the confusion. :)
LORAs and checkpoints are the results of the dreambooth process.
@@FrankTheTankGaming Thank you for the answer, but still a little bit dark forest to me :) So if I understand correctly- I can't get Lora without Dreambooth ? As that guy in my provided link explained, that to train Dreambooth it takes 1 hour or something and to train Lora it takes ~15 minutes. So I understood, that those processes are independent...
@@relaxation_ambience yeah be sure to watch my dreambooth video before going into that territory, because it will explain a lot better... Each type of training has advantages and disadvantages. You might end up with a favorite, but why not get great with all of them? :)
@@FrankTheTankGaming Thanx ! I go watching :)
make it simpler, embedding:ai influencer / hypernetwork: nsfw fanvue addon