Sign In

My tips to Lora Making (Standard sd models)

My tips to Lora Making (Standard sd models)

PREFACE:

I am not an expert but I have had many months of lora training under my belt by now and watched countless videos on the subject. I have found out one universal thing. People are all guessing on what things really do. I keep seeing ah yes this parameter does this but I'm not sure it actually does anything just keep it between this and that and its fine.

So instead of trying to explain everything I will focus on what I think is important. Even going to give you my lora training preset. A few tips on what you should alter if anything when training. To be clear these are not my original settings I found them and tweaked them a bit and here it is. I am going to assume you have basic lora training knowledge such as you can get through the menus alright and know where most the things you need will be at, as well as folder structure. If not brush up with a youtube video first.

TIPS:

Only several things you really need to change based on the amount of images and the time you wish to spend waiting for your lora to complete.

Image folder format:

As you can see from this poorly rendered image I have the number of steps (74) as the folder name and then the name of the character (yubin) followed by the type of training it will be for (girl) if you wanted a style you would put style instead of girl or woman. In order to get the number what I do is take the image count and divide it by 1500. So in this case I had 20 images so I took 1500/20 = 75 why is it 74 you ask? I normally either round down or lower the number by one. I don't know for sure if it helps with over training but I do it anyway. One thing to note if you have a number less then 10 always make it 10 it is the lowest training per image you should let it go down to.

(using a three letter word or something after the ex:(74_asd) I have tried and it doesn't seem to do anything for speed or quality)

EPOCHS: This may be controversial but you do not need a ton of these despite seeing what people put out. I often only use 1 or 2 epochs with great results. From what I understand its supposed to work on the generated imaged each epoch and make improvements however I often find the end result worst it over-trained the model making it hard to put your favorite character in new untrained outfits. Play around with it yourself and decide what you like.

Training Batch Size:

I normally keep this at 2 for all of my training but if you have lower ram you can use 1. From what I gather it just tells how many times to train at the same time. So it will cut your overall steps in half. It is faster but more GPU demanding. (some say that a lower number lets the training capture more detail but I have not seen any difference)

TRAINING MODEL:

You can use any model to train with as long as you read its usage rights ect.. Find one you really like and train with it nothing is stopping you. I personally like vanilla mix, some others like anylora and more.

Ill probably add more to this if requested if any experts out there and having a heart attack due to some of this information know that I don't claim to be a law on anything lora making this is just how I personally make them and it gives me good results when I do. I hope it helped a little with understanding.

3

Comments