How to compile a perfect dataset for your model

Image size matters

Our AI model will learn to replicate whatever it’s been trained on. To ensure that it generates only the highest quality artwork, you’ll need to use professional, high-resolution images in your dataset.

Here are the details:

• Minimum size: Your images need to be at least 1024 x 1024 pixels if they're square. If you’re using a images of a different shape, the total area needs to be the same as a 1024 x 1024 image. For example, a 16:9 image should be at least 1376 x 768 pixels.

• Don’t upscale: Don't try to cheat the system! Upscaling a smaller image to 1024 x 1024 won’t work. Your images need to be genuinely high-resolution even if it means re-exporting.

• Aspect ratio: Stick to a consistent aspect ratio (the relationship between width and height) of 16:9 or 9:16. While you technically can go up to 1:2 or 2:1, it’s best to avoid that for now.

Consistency is key

One of the most important things to remember when compiling your images is consistency. Your dataset should have a strong, unified style.

Your images should reflect your unique artistic style, not just a general technique. For example, don't upload a collection of images that are all line art or pop art. Instead, make sure they have your specific take on line art, your individual approach to pop art, or your unique style of digital painting.

Here’s an example of a perfectly consistent dataset. This artist has a clear signature style that is visible in every image.

Mixing styles is a big no-no!

If your work encompasses a number of different aesthetics, train a different model for each one. Even if it means excluding parts of your portfolio when training, the final model will have a stronger, more consistent style which is crucial for making your work commercially accessible on the marketplace.

You can't train a model to take the style from one set of images and the anatomy from another. It just won't work.

Here’s a dataset which lacks consistency. This artist is experimenting with different styles, which will confuse the AI model.

Variety is the spice of AI

While consistency in style is crucial, you need to demonstrate that style across a varietyof different subjects.

Think of it like this: the aim of your model is to learn how your style applies to new objects which you haven’t trained on. This means it needs a rich understanding of how your unique style applies to the world so it can generate more creative and diverse outputs to new objects.

Instead of showing your model 20 pictures of a cat in different poses it’s much better to show it 5 different animals, 5 different people, 5 different objects, and 5 different landscapes. Otherwise your model becomes fixated on cats and limits its commercial viability!

The Golden Rule: Aim for at least 2-3 images from each of these categories (the more, the merrier!):

• People

• Animals

• Nature

• Weather

• Food & Drink

• Travel & Places

• Objects

You can focus a bit more on one or two categories, like people, nature, or animals, but don't overdo it. Avoid uploading only people, only flowers, or only objects. Keep it balanced!

Why is this important? A diverse dataset helps your AI model understand different subjects, styles, and compositions. This leads to a more versatile and powerful model that can generate a wider range of creative outputs. Ultimately, this makes your model able to be used in a variety of circumstances which will attract more clients.

Think of your dataset as a visual representation of your worldview. By including diverse subjects, you're showing your model how you see the world — what you find beautiful, interesting, and worth capturing. This not only makes your model more well-rounded but also allows it to generate images that are infused with your unique perspective.

Here’s a great example of a diverse dataset: it includes people, nature, objects, and animals.

And here’s a dataset which lacks diversity. It only features portraits.

To sum up

To create a model, your dataset needs to meet these essential criteria:

• High resolution training images: all training images must be at least 1024x1024 pixels, and remember, the higher resolution the better!

• Consistent style: your unique artistic style should shine through in every single image.

• Diverse subjects: include at least 15 training images (and ideally more!) from a variety of categories, like people, animals, nature, and more.

By following these guidelines, you'll be well on your way to creating a successful model.

And finally, here’s an ideal dataset of 24 images