Skip to main content
All CollectionsGeneral GuidesHow to train custom styles
What Source Images Should We Use for Training?
What Source Images Should We Use for Training?

Source images are the basis of how Layer creates a custom model for your style, so it’s important to select the right ones for best results.

Updated over a week ago

The first step in creating a custom style is to gather 16-25 source images with lots of variations.

These source images will be the basis of how Layer creates a custom model to fit your style, so it’s important to select the right ones for best results. Here are some tips to help you make the best style possible!

Be Style Consistent

All reference images for a style must really be in the same style. Don't mix concept art and game-ready assets because you'll get inconsistent results. Make sure that all the images are in a consistent style. Use the same game and the same context.

Use Variety of Images (different poses, colours, profiles)

For non-characters, add different types of trees, rocks, bushes, flowers, etc... whatever it is! Be sure to diversify. For characters, add images of your characters in various environments, wearing different clothes, while they're sitting, standing, close-ups. holding different things and doing various activities.

To Generate New Characters, Vary Your Uploads

If you want to create different poses of a specific character, you should upload different images of a specific character. Or you can upload different characters to generate new characters in the same style!

For Character Poses, Stick to One Character

For characters poses, especially if detailed, try to use at least 25 images of the character itself with a combination of headshots, closeups, etc. This will result in better face generations. For example, do not train your style with a combination of character + trees.

For Non-Characters, Mix Up The Objects

For non-characters, try to upload 15-20 images. Unlike characters, they can all be different things e.g. chest, tree, house etc.

Don't Repeat Yourself

Even if you are going for a highly specific style, try to add some variety.

Use Source Files Only

Images should depict one thing only. Only upload the source files of your artwork. Make sure the images are unobstructed and don't upload screenshots.

Make Sure The Images Are Clear

Upload high-resolution images with no distractions. If you upload an image with a border, your Forges will have it too.

Upload Square Images:

In order for best learnings, you need to upload square images, they can be at any size as long as the ratio is 1:1. If you have non-square images you need to adjust them to be square. This is especially important when training characters as you'd want to make sure proportions of the character don't change when producing outputs.

Removing backgrounds:

If your input assets have a particular background, don't worry about removing them unless it's important for you. Layer allows you to remove backgrounds on canvas before exporting so you can use that functionality. If you have particular background in mind to achieve in your outputs, you can stick to it in the training for consistently same backgrounds in the outputs. For example: if you like to produce with white background make sure your uploads always have white backgrounds.


Uploading images respecting the upper limit is critical, especially for detailed items like characters, buildings, backgrounds, you may get away with less for less detailed items. Varying your uploads making sure you are not repeating yourself are equally critical to ensure a successful training.

And that's it! Next in the style creation process, you will have to write captions for each image. You can learn more about this here:

Did this answer your question?