<aside> 🇬🇧 This post has been translated into English by AI, keep in mind that it may have some errors

</aside>

<aside> ☝ Everything shown here is just trial and error, and things I have done in specific cases. I still have to experiment a lot, and it could be that doing this "excessively" could damage your LORA.

</aside>

The basics:

The important thing is not quantity but quality.

It is often recommended to make LORA with 20 images, but if you have a very good dataset you can do it with up to 10-15 images. Remember that the main thing is to have variety of clothing, angles, positions, expressions. Of course, if you can choose between a dataset with many images and one with few, choose one with many. But if you have no other choice, follow the advice in this guide.

Low quality and quantity? Don't worry, use the following tricks:

Haz el dataset manual

The most practical advice I can give you is to rely less on collaborators who do "scraping" of images from the internet, as they are counterproductive when we want to make a specific character that has few fanarts. Of course, this requires much more work, but it is ultimately more worthwhile.

The "Hidden" Sources

Websites where you can find many images of anime characters:

• Pinterest: Here, for example, I found images of Kanna from 20th century boys for my model. It's a very good website that has sometimes given me even more images than Google.
• Gelbooru: The booru par excellence since the images that are usually uploaded here have a minimum quality.
• Pixiv: (Requires creating an account, its quality varies greatly, it is also recommended that you activate the option to view +18 images)
• Twitter (Many of the images on Twitter do not appear on Google indexed, so having an account to search for images and fanarts should be essential)