Bodytape - 500 followers special
Hello to all my followers and to others,
Thank you for your support, this LoRA is dedicated to you!
I would like to show you my appreciation and affection with this smoking hot Bodytape LoRA inspired by Black Tape Project.
AAAAND, A LoRA MAKING TUTORIAL!
Check my kofi page and my tutorial here: https://ko-fi.com/post/LoRA-Making-Tutorial-R6R3JEC2M
Please consider supporting me through a donation if you liked it or commissioning me for a LoRA!
With that being said, onwards to the bodytape!
Description
FAQ
Comments (13)
Looks great! I'l add an review later!
Thank you, I will be looking forward to your review!
Big fan of your Lora's, and about to follow your fantastic looking training guide to make my first!
When do you use drawn Vs RL images, and how much of an affect does that have on being able to use the Lora for realistic/anime models?
I noticed you make a point of prompting the subject you're trying to train, but from what I read about Textural Inversions, is that you're meant to ommit the subject so it learns what to reproduce. Is this a difference between a TI and a Lora? Do you always prompt what you're training?
First of all, thank you for being a fan of my work and for appreciating the details that come into it.
I do not know if you have fully digested my tutorial yet, but I believe that the answer to your question lies in it.
I suggest you read it and ponder upon it first, and then ask your question as it is impossible for me to give you as detailed an answer here in this comment section, as I have in the tutorial.
These are different!
Comments on your tutorial
Definitely a good running start for anyone who is new to making LoRAs. If you don't do those things, you aren't going to get good results, no matter what else you do. I am certain of that much.
I guess from my pov, it's what comes after this that is eluding me.
- Do I need to crop to 512x512?
- Do I use a Colab notebook for training?
- Where do I put my folder of images and captions? And how do I access that from Colab? (I've only barely used Colab, since I mostly work on my own Mac.)
- What do I enter for learning rate?
- How many steps?
- I already know that I want to train on NAI if I want my LoRA to be usable with pretty much any anime-based model, although a lot of makers don't really follow that I see. And how do I access the NAI model? I believe that Colab only supports the Diffusers library, and not ckpt/safetensors? But NAI isn't on HuggingFace.
Those are the questions that I really want to know the answers to. And so far, pretty much everything I've read is like, "Open this Colab notebook and do what it says." But that notebook makes a lot of assumptions, like… that you know what you're doing, lol. And I definitely don't.
I'm guessing this is just an early draft, or that you have a Part 2 coming with this kind of information? I'll be looking forward to that especially, but like I started my comments with, I understand the importance of starting off on the right foot. It's essential for a desirable outcome, no question. So reading your recommendations for that is also very useful. Thanks!
My dear @dita,
I am ever so grateful for your comments.
This tutorial is not a draft, it really is a part 1 for people who want to get inti LoRA training but are afraid of technicalities.
Also, believe me when I tell you that THAT is the most important part of LoRA training.
Now, to answer your questions...
No you don't need to crop the images, otherwise I would have told you so in the tutorial.
Learning rate is synonymous to epoch and I have addressed that in the tutorial.
Steps are a consequence of the learning rate.
Beko's Ikemen World was trained on Anything 3.0, so NAI isn't necessary
Collab is poopoo, I manage to get it to work once every full moon, that's not what I use
@EDG I assumed either a draft or a part 1, since it doesn't really get into the nuts and bolts. But yes, I definitely believe you that this is the most important part. It's like in culinary school, the first thing you learn is food safety. Then knifing. Then you get to touch the food.
You didn't really have to answer my questions here, but thanks. ^_^
Really, I think Anything 3 is probably just as good a candidate as NAI, since so many models are derived from it, too. Even AOM is derived from Anything 3, I believe.
What do you use for training, if not Colab? You train on your own system? As a Mac user, with an AMD GPU, and only 4gb VRAM, that's not really an option for me. I'm loathe to do Colab, but working on my own system is pretty much out of the question. I'm just happy that I am (somehow) able to generate images at 1024 x 1536 and possibly higher. When I first started with SD, 768x768 was the limit before I hit out-of-memory errors.
I understand your situation and truly I wish that collab was more reliable in order to include people such as yourself amongst the ranks of model makers...
Here's a non exhaustive list of things that happened to me while using collab and why I am not recommending it:
- Disconnected during training (if you are training for free then Google will randomly unplug you from time to time just to remind you to pay, RIP your training)
- Endless bugs (I had to get inside the code every time and make changes in order for it to work)
- Random changes (the person making the notebook changes stuff and recently broke the notebook so hard it's not even funny, it can't even install xformers properly for training anymore)
This collab solution is not only unreliable, when it works, it's slow...
At any rate this situation is quite problematic.
Be sure that if collab gets patched, I'll let you know
@EDG Oh, I know slow. It takes me about 90 minutes to do a hires fix on a 512x768 up to 1024x1536, at 15 steps. Compared to that, Colab has to be ffffffast. I saw where someone said that, for about US$1 on a paid upgrade, you can train a LoRA in about an hour. I don't know how accurate that is, however.
I always make a copy of the notebook whenever I find one that I want to try playing with. That means I don't get any of those pesky updates. But it also means I don't get any of those updates unless I manually make a new copy.
As for altering code… sigh I was a computer science major when I was in school—25 years ago. At that time, my school was still teaching everything in, of all things, Pascal. They were just starting to migrate to C when I left in 1998. Heck, they might even be up to Perl now. Long story short, I never learned Python, nor did I ever have a need to. I've tried a couple times to pick it up, but without a real need to use it until now, I've always dropped it again. Same with Rust and Ruby. Of course, the big thing about learning a language isn't the language itself, but the libraries. I can pick up the syntax in no time. But I don't know the first thing about all the packages, even the ones that are a part of the language itself—string manipulation, networking, reading/writing files, that kind of thing. It takes a while to learn all those.
Maybe I'll just stick with begging/commissioning others to make them for me.
@dita I sincerely wish collab were more reliable!
@EDG I know there are other services out there like Colab. I think I've got at least one of them saved in my Raindrop bookmarks. Never tried any of them, and they're all paid afaik. But no matter where you borrow/rent compute time, you still have the lack of reliability of the notebooks. Unless you're knowledgeable enough about ML and Stable Diffusion to write your own. Which I'm not and never was.
I've gotten some good results with this on anime models too, just need to use a lower weight on the lora.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.