Anyone here made their own Lora and can share some experiences?
Hi there,
I'm trying to get (more or less) historically accurate images from the early and high middle ages, but none of the models seems to have a grasp of what "maille armor" or "bucket helmets" are and I either get complete garbage or fantasy armor that mostly resembles the early modern period (the stereotypical shining knight armor).
I assume a Lora, trained on images of the armor and weapons I'd like to include, could fix this problem. I found some neat tutorials for making Lora's and think I'll give it a shot.
Do any of you have experience in making these kinds of style Lora's? What should I take care of? I will be manually downloading images that fit my aestethic and manually tag them - how many images should I use?
Any input here is highly appreciated.
I too have just started on my LORA-making journey, and I too am interested in ahem specialist apparel. My experience is that most LORA are made by non-enthusiasts who don't necessarily know how enthusiasts refer to things, and to some degree non-enthusiasts want visual variety so they can churn out "dwarf in armour" and "elf in armour" prompts and get things that actually look different. That is fine for most people, who just want to have some nice pictures to go along with their D&D campaign or whatever. But if you are a discerning connoisseur then yes, you kinda do need to roll up your sleeves and make it yourself.
There are some guides out there for LORA making - As ever, they are a mix of helpful and not helpful, and you are going to end up having to work things out yourself. You are definitely going to end up wasting a lot of compute time on LORAs that just fail. That's part of the process. You are going to see a lot of parameters which you don't understand and that have seemingly absurd values.
Before I jump into the rest of this - I strongly advise you to start out with LORAs that do one specific thing, and only that thing. So, a LORA just for bucket helmets, using just images of bucket helmets. You can make more complex LORA, but holy crap this is a complex process with a lot of moving parts, so start out with just one thing that you can easily tell if its working and how well.
It is good to hear that you are mentally prepared to manually tag your own images because this is utterly essential and you need to do a very thorough job of it. When training stuff the rule is "garbage in, garbage out" and there is no shortcut here. I honestly haven't found a good tagging methodology yet, but the advice I'm trying to follow is that you want to tag anything that you DON'T want included in the training term, and don't tag things that you DO want included.
So, you have a picture of some reenactor in brigandine - You tag it as "brigandine", but you don't tag "rivets" or "visible plates". You would tag "black leather" because brigandine could be any colour, so you call out the colour so the AI can see that the colour is separate from the armour design. You would also tag the trousers, the helmet, the person wearing the armour, the background, the lighting, the time of day; and then also add in "good quality" or "cropped" or other terms about the photograph itself. This sounds like overkill, but if you don't do it right then the LORA will do all kinds of weird things that you didn't intend.
To give an example - On the first run of my first LORA I was actually kinda shocked that I was getting good results for the garment that I had trained on... But the LORA was also changing the skin tones and the white balance in the image. The training data was skewed towards very warm light and tanned skin tones, and I hadn't tagged that, so the result was the AI also associated the training terms with olive skin and incandescent light and they couldn't easily be separated. I had to go back, reprocess the images, retag them and then come back again.
Which brings me on to the images - You want the largest, highest quality images you can find. You want a range from long shots to close ups, but don't use anything too close up because SD needs to see how the armour relates to the rest of the figure.
You don't need to train on huge sets, but I strongly advise that you grab yourself 100+ images and then aggressively prune that collection down to around 30-50 of the best quality images. You should run all of them through some denoising, and for almost all of them you should fiddle with the colour balance. You don't need to perfectly match colours or anything, but when you see that the light is a bit orange or the image is dark, just change the levels a bit so they are more neutral. You also want to manually crop the images, and (AFAIK) they do have to be 1:1 squares which often means having to crop figures.
As for the actual LORA settings - Don't ask me what they do, I don't know. I have been kinda kludging together suggestions from different guides and just seeing what happens. I know for certain that my LORAs are training too fast and typically are burning out by about epoch 6 or 7, but I have no idea how to fix that.
I would recommend that while you are learning you set the trainer to save every epoch and print a sample every epoch. This is because Kohya is fucking twitchy and can sometimes just hang during an 8 hour run, and also so you can monitor the training in real time and see what is happening. I've never had a training session that actually needed to run to the end, but the sample images showed training, then good fit, then blurring and overfitting, and I quit out at that point. When you are saving every epoch you of course keep every version to test out, even if there is a crash or you quit, but you can also do a nice X/Y plot of every version of the LORA next to each other and find where the sweet spot is.
Also, if a LORA is just for personal use you don't necessarily need perfection to get results that will work for you. The standard for most LORA is that they are very transparent and compose well with other LORA and work on lots of models. That's a lot of work man, and if you aren't using it that way you won't even appreciate it all.
Instead, I have been using a combination of control nets, latent couples and composable LORA to apply my mediocre quality outputs to specific parts of specific images. It's a faff, but you can generate a nice knightly figure, freeze his pose, mask off just the torso, specify brigandine, then generate an image that will probably be very good even with amateurish LORA creations. And that way you don't have to worry too much about your LORA melting people's faces and turning their hair pink.
wow, thank you for the detailed answer! I think you have saved me a couple of wasted hours right there :)
I think I'll start with a Lora on Bucket Helmets (which seems to be the most straight forward of the things I have in mind) and see how that goes. Your detailed comment on tagging was especially helpful, before that I thought I would just tag "bucket helmet", but not it seems that I'll need to be a lot more thorought than that.
Yeah it's crazy frustrating that the most important part of your exciting AI endeavour is tediously tagging images. But it really is important. And, to be fair, it doesn't take all that long. Get your images together, processed and cropped, then just sit and blitz the tags out in under an hour.
I read guides that said auto-tagging stuff is absolutely fine, but remember that a LOT of user generated SD stuff is anime/manga based, which doesn't have the problems with lighting/exposure, and the figures all have nice clean outlines.
Nice answer! Is there a number of concepts, to your knowledge/experience, where a LORA no longer works? For instance, if I want to make a model that understands all car brands and types, assuming that the base model doesn't of course, would a LORA still be sensible here?
Most LORAs I find have a more specialized narrow focus, and I don't know if I would just start with multiple LORAs dealing with individual concepts (a LORA for a "1931 Minerva 8 AL Rollston Convertible Sedan", a LORA for a "Maybach SW 42, 1939", etc.) or if I should try and generate one LORA to rule know them all...
It would definitely be unwieldy to make LORAs that were so obsessively granular, but it comes down to how specific you want the output to be. It's probably pretty easy to make a Porsche LORA that will spit out credible Porsche silhouettes, but if you want accurate and detailed 911 Turbo S models you have to train it just on those. More accurate means less flexible; you have to pick your poison.
If you want an omni-functional thing then you really want to make your own checkpoint... But LORA are still a potentially good way to approach that because you can merge LORAs into a model, and you can merge LORAs into each other. I have no idea how you do that, but the functionality is there.
My plan is (eventually) to do exactly that with my own niche area of interest, build a model but do it in discrete chunks. So start out with one specific LORA for one specific garment that I have a particular interest in, get that to work well, then make others, then when I have enough that it's tiresome to sort through start merging them together.
That said... I am not an expert, not even slightly.
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.