I encourage you to judge that yourself! The previews above should give you a rough idea of what level of quality you can expect.

When i started making AkiLora, I was convinced that automating the process would inevitably decrease the quality of generations. Afterall, many SD gurus claim that making a good Lora requires hard work.

The very first Lora that was officially made with AkiLora - Valeera Sanguinar - completely surprised me. I uploaded it to Civitai and found that It nearly doubled the average upvote rate of comparable Loras!

When compared to hand-made Loras, Aki-made Loras tend to be:

Superior in character accuracy

Superior in preserving high-quality artstyles of base models

Superior in working with other Loras

Inferior in accuracy of clothes

Larger in file sizes

To make a good Lora, you need either a small dataset with really high quality images, or a large dataset of low quality images. Neither approach is strictly superior.

Try booting up any third-person 3D video game and screenshot the main character 30 times from all angles. You can do the same with movies. If you combine such dataset with several hours of fine-tuning your model, you just might get a really accurate Lora. I've been there. I've seen people be there. Note that "accurate" does not equal "good". This approach has two major issues:

• Your generations will start looking like a blend of your dataset's style and your base model's style - which, most of the time, results in an ugly and deformed frankenstain. This tends to happen even if both styles are relatively similar. In short - your overall image quality will take a hit.


• Your model will struggle to make your subject do specific things or be represented differently. It will be difficult or buggy for the model to make the subject, for example "drink a coffee in a busy café" or wear "cat ears headphones".

You can also go the other way. Visit your favorite art site, search for your subject, and download a lot of pictures. At least several hundred. Next, you need to filter them out. Leave only art that has decent (but not necessarily high!) quality and includes only your character (or crop out others). You should still end up with at least 50 images, but I recommend much more. You can go for less, but at that point you might as well go with approach #1, you will get better results. The whole idea is to get a large dataset that will overshadow all of approach #1's problems with brute force - And that is it's biggest advantage. If the Lora is done properly, your model will be much more flexible. The subject will "drink a coffee in a busy café" while wearing "cat ears headphones" with little prompting. Other Loras will work well with it. The high-quality style of your base model will remain mostly intact - much more so than with approach #1. But there's a drawback - accuracy. Artists tend to take shortcuts while making art, and this confuses AI.

• If your subject has intricate patterns, they will be less accurate with the original. This is predominantly visible on custom clothes, but tends to have little to no effect on subject's biological or "I'm never taking this hair ornament off" features.

I personally prefer approach #2.

And so I made a script that only takes 2 inputs - a subject and a model. It downloads images, filters them, fixes them, rents a remote graphics card and trains a model. I was happy with the results, so I turned it into AkiLora so the world can use it too.

Image boards! Stuff like Danbooru.

Unfortunately, Akilora does not support this feature right now.

You can't. That's by design.

What values you put in depends on your dataset, and are pretty much unique every time.

I did my best to create an algorithm that uses math to calculate best possible settings for any dataset. Messing with these values would break all of that, and at that point it's better if you do the whole thing yourself.

AkiLora aims to be a simple tool for people who are not tech enthusiasts like me and you and would prefer to have someone else define things like "epochs", "mixed_precision" or "gradient_checkpointing" for them.

If there's demand, I'm open to adding these features in the future.

It usually takes less than an hour. In rare cases (vaporeon-tier colossal datasets) it can take more time.

Akilora is much faster than you would expect for such a service. Unlike most cloud providers, Akilora exclusively uses uninterruptable GPU providers. This is significantly more expansive for us, but allows fast and reliable generations.

If your generation does not complete within 2 hours, the algorithm will automatically cancel it and try from scratch on another GPU, somewhere else in the world.

Still, faster is not better here. Beware of services offering overly fast generations. They may have better hardware, but It's more probable that they are undertraining models.

For non-flux character generations, $1 is charged, regardless of steps.

For any flux generations, at 10 000 steps, $5 is charged.

For all other generations, $4 is charged, regardless of steps (by default, "Any tags" generations have 4 times as many steps as "Character" generations).

TLDR: Yes, as long as you credit akilora.com as the creator!

I will share some generations on Civit Ai, unless users opt out. Please note that if for any reason your generation is free, you cannot opt out.

Even if you opt out, It's perfectly okay to share the model yourself, anywhere. Just bear in mind that according to Akilora's ToS, you have to attribute akilora.com as the creator of this Lora.

I don't guarentee that I will share your model even if you don't opt out, so if you want the world to see it, definitely share it yourself!

Also, even if you did opt out, someone else may generate the same character without opting out, which means you might see a copy of your model shared anyway.