All Collections
Styles
How to build a good style dataset?
How to build a good style dataset?
Updated over a week ago

💡A dataset is a collection of images you want the AI to learn a style from

  • The quality of the images should be as good as possible. If pixelation, noise, image artifacts, or watermarks are present in the training images, chances are high the AI will pick up on those.

  • Images should have a consistent and cohesive style. One way to approach this is to consider whether all the images fit into the same overall theme or world, or asking: what do the images have in common?

  • The bigger the variety of what subjects the images depict is, the more flexible the style will be to prompt with for things that aren’t in the dataset → vice versa, if the dataset contains only portraits, it will be hard to generate images with it that aren’t portraits. Even further, if the dataset only contains portraits of women, it will be hard to generate portraits of men with the style.

🖼️ Example 1:

In this dataset all images feature the same color scheme and mostly flat shapes, so those would be the style elements the AI would pick up on the strongest. Each image has a different subject, so it won’t associate the style with a specific subject.

→ Images generated with a style made with a dataset containing the 5 images above:

The prompts are:

  1. “a beautiful daisy flower, 8k, high quality”

  2. “portrait of a beautiful woman with long hair, masterpiece”

  3. “an elegant car on a highway, 8k, high quality”

  • For a very versatile style a good ratio to use is 70% people, 20% landscapes, and 10% animals and objects, but in the end it depends on what you want the AI to be able to generate.

  • The absolute minimum is 5 images but if you want to be on the safer side 10-30 would be better. The smaller the dataset is, the more important good curation will be. Above 12 images there’s usually more leeway.

💡 Images need to be in 1:1 ratio and 512x512px

🖼️ Example 2:

The common element in these images is the isometric view and all of them depict a building, so that’s what the AI will focus on and it can be hard to prompt for images with other subjects & perspectives. Because beige dominates 4 of 5 images, the results generated will be biased towards that.

→ Images generated with a style made from the 5 images above:

The prompts are:

  1. “a red hospital building, emergency station”

  2. “a greenhouse made of glass, victorian, lots of flowers”

  3. “a pyramid, lots of plants and flowers, futuristic, highly detailed, pink yellow and blue, high contrast”

Did this answer your question?