How to create your own AI model

Artist avatar

Varya Plusheva

Head of Community and Customer Operations

Apr 15, 2024
How to create your own AI model

On, you can create your AI model by uploading images that belong to you. Your artwork is the most essential ingredient to create a model, so let's ensure it meets all the requirements. Go to the Models page on your workspace and press the Create New Model button. You need to upload at least ten images. Those images can be in any file format. But make sure that they are not too small and that the minimum size of the shortest side of the image is 512px. The bigger the pictures you upload, the better. No worries if you don't have ten images in one set style. You can try to use five or six instead. But the more images you upload, the better the model will be.

A good model can create new images in a very similar style to the uploaded ones. Two things will affect the quality of the model: 1) the consistency of the uploaded images and 2) the textual descriptions of the images you write.

Let's talk about consistency first. A consistent data set (as we call it) is one where

  1. all images are created with the same technique
  2. colour palette
  3. it should look like a series of artworks or like brand illustrations made within the brand guide by the same illustrator
  4. if you upload ten illustrations of dogs, a model will be capable of creating new images of dogs in the set style
  5. when you upload watercolor landscapes, a model can create new landscapes in the set style.
  6. try using scans of your images instead of the photographs
  7. AI will learn from the image you upload, so if you have frames or fingers holding a picture on the image you upload, your model will generate images with frames and fingers

Textual descriptions: You need to provide textual descriptions of the images you have uploaded. When your model is ready, make sure you use prompts similar to your textual descriptions. For example, uploading images of apples in a set style can prompt your model with other fruits. However, it would be challenging for a model to create portraits if it is trained on still-life images. The model will still be able to generate an image, but the style might be a bit different from the original artwork.

provide textual descriptions that are

  1. not referring to a particular medium (don't say a watercolor illustration of an apple, write an apple)
  2. don't describe a style pattern that you want a model to learn
  3. describe just objects that you see on the image

If you don't follow the rules above, you can still create a model. Some pro artists intentionally avoid following those guidances to get unexpected results. We encourage you to experiment with different approaches to find the best one.

  • getting started