<aside> 💡 Visit our express guide for optimal datasets

</aside>

There are 2 ways to reach great results on trained models.

• Improving your prompt
• Improving your dataset

Most people focus on the first point, but it is a rookie mistake. No matter how good your prompt is, if the dataset you input is not good, the results will not come out good.

So here are the best practices for picking your datasets and improving their results :

• Format : your subject must be in a square
• Quantity over quality except for styles
• Diversity over similarity except for specific consistency

Format

To train models we must input images in 512x512 format.

• If the image’s resolution is higher, we downscale it automatically
• If the image has another format, we crop it at the center. This means that any mobile format or wide image will be cropped. If you’re looking to train a person and their face is outside the center, it will be cropped.

We advise using this link to quickly reposition and crop series of images.