Stable Diffusion

I just wanted to come back and update my earlier post because HOLY SHIT I have radically improved my LORAs over the past couple of days.

Firstly; I was flat out wrong about the need to heavily tag, at least for things like garments. The guide I was following talked about styles and objects as the two types of LORA. I thought I was doing one type when actually I should have been treating it like the other. I tore the tags apart so that almost my whole training set was just trained on the single key concept, some had one or two extra terms at most. So, Mr Shiimiish can tag his helmets in a much more chill style.

Secondly; at least in my experience you need to twiddle with the Kohya settings just slightly to get genuinely good results. I was getting burned out LORAs by generatiton 6 or 7 before, but I turned down the learning rate to 0.00005 (half the default) with 7 repeats per epoch (instead of the laughably high 40) and it's much much better now. The jumps between each epoch are much smaller and you can get a much more granular picture of what is actually going on. I also turned the rank, network and module dropouts to 0.1. Kohya says they are the minimum recommended values, but by default they are set to zero. I have no idea what those do, but I am definitely getting better results. I'll do a codeblock with my full settings at the end.

Finally; and very very very importantly - CHECK THE FUCKING FOLDERS. If you, like me, got a bad LORA, reloaded the old config to change it around a bit then set it running again you will discover that Kohya kept the old training set, and if you changed the number of repeats the old and the new training set will now be in differently named folders but BOTH will be used in the new run, so it'll just fuck up again and also take much much longer on your second run.

I have also been told to use regularization images - Not to download them, to make some then use them. So, go to your SD, plug in the model you are training the LORA on, give a prompt that will create "things without the LORA item on them". So, if you are training for dudes in chainmail, put in "30 year old athletic man standing up". Have SD generate like 500 pictures worth of that (I have no idea why that many, that's what I was told). The idea is that you have your training data of dudes in armour, and then you have the regularization of similarly constructed dudes not in armour, and SD will look back and forth to help it figure out what you are trying to train it on. You need to use the model you are training on to do the rendering though, so make your own and don't download other people's even if it seems similar. Just be patient and let it run.

Since I mention models - It probably goes without saying, but don't use the base SD1.5 model unless you actually intend to generate images with it, because it is... It's not wonderful. It's alright, but get something that's more appropriate to your needs. There's plenty of stuff that's been trained on LOTR and GOT that are at least better at understanding that a cuirass isn't a type of dress shirt. If you are doing anime, use an anime model, etc etc.

Getting all this stuff right has radically improved the LORAs. They are significantly more transparent; the styling in the original image will change somewhat unless you prompt against it (it's an inevitable result of imperfectly balanced training data, it's why so many LORAs make your people look unexpectedly Japanese) but they don't make people's faces melt. With the last LORA run I did today I was testing epoch 29 at CFG11 and still getting good clean images with no distortion. Previously I would be running epoch 3 or 4 at CFG3 to not get a Daliesque nightmare. Huge improvement all around, and I no longer have to take the earliest epoch that reproduces the item, there's a big range to test and see which best preserves detail and plays well with others.

Here's the Kohya settings that were actually successful - You can't quite just copy and paste because you will need to set up your own folders correctly yourself, and choose your model and sample prompt and all that. But you can at least run your eye down the values and copy those across. I'd say you want to run 30ish epochs from this, based on 30 to 50 good pictures. One LORA I ran today was good at about 20, the other at about 30. That might take a while, apologies about that, but I am running on a Shadow pc with an A4500 20GB, and it turned into about 1hr 45 to do 30 epochs which is pretty reasonable.

I figured I'd at least post this up so any future garment enthusiasts could at least learn a bit from my monkey-at-a-typewriter approach.