santa hat
deerdeer nosedeer glow
Sign In

How to train character lora with super low data (Illustrious/Pony)

How to train character lora with super low data (Illustrious/Pony)

"This is Layna. We trained her wrong on purpose, as a joke".

INTRO

Is it possible to make character lora with low data? YES

How low? Well, very low; in this tutorial I'm going to make a lora with ONE image.

But first let's address the elephant in the room: the quality of a lora made with one image is always going to be crappy. Obviously, it is never going to be as good, accurate, or flexible as lora made with a hundred pictures, not even as good, accurate, or flexible as a lora made with 10 pictures.

This is a last resort sort of thing, something you do when you are so desperate for the character that there is no other option. Your waifu is so obscure no one ever made fanart of her, or the game came out just yesterday, but you want lewds today.

Also, this will only work on anime-ish or western comic-ish characters; it won't work if what you want to replicate has a very strong distinctive style, like... I don't know, Hazbin Hotel or Jimmy Neutron.

 

SAMPLES

I'll show the results of this first, with the only image used for training on this tutorial in the center and Illustrious (top two gens) and Pony (bottom gens). So you can decide if it's worth your time to keep reading or not.

I'm the first person aware the quality is below average lora, but, eh, beggars...

Not good enough for you? Can stop reading here. Desperate enough that this looks good enough? Keep going.

ILLUSTRIOUS:

Layna (Soul Nomad) XL - V1 | Illustrious LoRA | Civitai

PONY:

Layna (Soul Nomad) XL - V1 | Stable Diffusion LoRA | Civitai

This one is particularly weak because we are using a very low-quality image as a base.

The results are improved quite a bit with better raw material and with more effort.

Here you have a couple of lora made using these methods but with a bit more effort and better raw images, but also made with only one image.

Nil (Eiyuden Chronicle) XL [Pony + Illustrious] - Illustrious Ver | Illustrious LoRA | Civitai

Carrie (Eiyuden Chronicle) XL [Pony + Illustrious] - Illustrious Ver | Illustrious LoRA | Civitai

BEFORE STARTING

Before starting, I'm going to assume you know how to train a basic lora character. How to add images, how to tag, how to set the settings for training, etc.

If not, you should start by checking that before attempting this.

CHARACTER

This is Layna.

Layna is a character from the game Soul Nomad (Soul Cradle in some places) that was released back in 2007. There is a grand total of four images of Layna on the internet (that I know of, or were the last time I checked).

We could do this with the portrait; we could also use all four images; there is really no reason to force using just one. But let's pick the best one alone for the sake of this tutorial. We are going to use the first image only so you can see it's doable with one image alone (DOABLE doesn't mean good).

It's a low-res image, is sketchy, and that is going to show, but it at least portrays the character well enough.

 

METHOD ONE (Fast): Illustrious only.

Technically this method can be used with Pony models as well, BUT I tested it myself, and it simply does not work very well. It makes the lora, but the results are as stiff as a statue, and it's almost impossible to make the character move unless you lower the str enough, but by that point it loses all of the likeness. So, I'd recommend this method for ILLUSTRIOUS only.

 

Step 1: Pick your unique image and open it with an image editor.

Step 2: Proceed to crop the image around the waist and save the image.

Step 3: Repeat the process, but this time crop up to the neck.

So in the end you should have THREE images. The original one plus two cropped versions like these.

Step 4: Manually tag each one of the three images because you don't want it to learn anything more than necessary.

Image 1: 1girl, short hair, black hair, hair ornament, earrings, jewelry, black hair, high collar, medallion, harness, belt, blue dress, brown coat, rapier, skirt

Image 2: 1girl, short hair, black hair, hair ornament, earrings, jewelry, black hair, high collar, medallion, harness, belt, blue dress, brown coat

Image 3: 1girl, short hair, black hair, hair ornament, earrings, jewelry, black hair, high collar

Step 5: Copy those 3 images, with their captions files, and duplicate them a couple of times.

So you will have each one of those 3 images 3 times, for a grand total of 9 images with their 9 caption files.

 Step 6: Proceed to set up the training values. Aim for about 1200 steps. Not going to go much into detail here because in this occasion I am going to set all the data for sharing, so you can see exactly the steps, the repeats, and even download the data on the model page, on the right.

Step 7: Start the training and wait until it's done.

Step 8: Profit

Layna (Soul Nomad) XL - V1 | Illustrious LoRA | Civitai

METHOD TWO (Slow one) Pony/Illustrious

This one is the best method but also can be the worst method. Rely only on this system if you really LOVE the character, because otherwise the results will be worse than the previous one.

The idea for this method is simple. For training a lora you need images, but your character does not have images, so... you only need to create them.

And the way to do such a thing is to use a 1.5 Lora to generate the images yourself.

So, basically, you generate LOTS of images of the character in different poses, pick the 15ish best ones, and then fix the images the best you can to make them the most consistent and accurate possible.

You need to create a lot and pick the ones that most closely resemble the character and use that for the training.

Ideally you should also go around and fix the hands because if you use them without fixing them, your final lora will have awful hands.

 

It is not a hard job, but it's a long, tedious one because you need to make lots of gens that look close enough to be usable. 1.5 gens were not always as accurate as one would like. Then fix them the better you can.

Theoretically speaking, you can do this by baking an Illustrious lora with method one instead of using a 1.5 lora.

I suggest using 1.5 because, despite having poor quality generations compared to illustrious, it is easier to retain part of the character's original style. So when you train the final XL model, it will still hold some of the original style. If you do it with illustrious, which is perfectly viable, the images you will use for the final lora train will be mostly in your illustrious model's natural style, so on the final XL model, the original style will be entirely lost. The design will still be the same, of course, the same elements in the same places, but it will lose a bit of the flavor.

Now, obviously this depends on the character, of course. Hypothetically speaking, if your character is from a notable artist, you should be able to make the illus lora and generate plenty of nice images in that same style to use it as a base for your final lora. In that particular case, doing the generations with an Illustrious would be better than using a 1.5.

Also, truth be told, sometimes the final result will be washed away anyway, even giving your best efforts using 1.5, so using 1.5 instead of illustrious for generating the data in those cases was kind of a waste of time.

So it's a matter of preference mostly. I suggested 1.5 because of that, but doing it with illustrious is perfectly viable too. 

For the sake of this tutorial, I considered the worst-case scenario. Let's say the 1.5 lora for the character you want does not exist. If that's the case, you only need to do one yourself. SD1.5 is way more lenient when it comes to required images for the training, so only with the three images we made earlier should you be able to do one functional enough to generate the images for making the final XL model.

I made, back in the day, a 1.5 Layna lora, but let's pretend it does not exist, so I made a new one using that very same image we used earlier only.

Layna 1.5 (tutorial purpose only) - V1 | Stable Diffusion LoRA | Civitai

I generated a bunch of images, picked a few (didn't go the extra mile of fixing them, sorry, not like the character that much), and baked a basic pony lora with the generated images.

I also left all the data intact so you can check out the details of the training.

These models tend to be rather stiff, so using them at a str of 0.8 or even 0.7 is advisable. At 1 they will barely move.

Layna (Soul Nomad) XL - V1 | Stable Diffusion LoRA | Civitai

METHOD THREE (NOT RECOMMENDED)

I will mention it, only for the sake of mentioning it, but it is not something I would recommend.

You can always pick your one single image, give it an unholy amount of repeats, and train the lora only with that. You will get a lora...but it will most likely not be functional at all.

 

TL;DR

If you want to make a lora with one image, you have two options.

ONE: You use the same image 9 times but cropped at different sizes. This only works with illustrious.

TWO: You make a 1.5 lora, generate images, fix them, and use those to train the XL lora.

 

Closing up

"These look like sh***!" Well, yes, I warned you at the beginning. The only way of getting a good lora with one image is by using METHOD TWO and putting on it lots of love, generating lots of good images with the 1.5 lora then using them to make the final one.

I'm giving you an option here in case you fancy some character and you just want to create some fast, quick images without putting much effort into it.

All these Illustrious models were done using Method One. The pony versions were made back in the day using Method TWO.

Elektra (Eiyuden Chronicle) XL - Illustrious Ver | Illustrious LoRA | Civitai

Mio (Eiyuden Chronicle) XL - Illustrious Ver | Illustrious LoRA | Civitai

Carrie (Eiyuden Chronicle) XL [Pony + Illustrious] - Illustrious Ver | Illustrious LoRA | Civitai

Milana (Eiyuden Chronicle) XL [Pony + Illustrious] - Illustrious Ver | Illustrious LoRA | Civitai

Then again, they are not as good as a lora made with 60 or 100 images, but for making some quick gens for reasons, they can do the job.

29

Comments