How to fine-tune on Ilus AI platform

What is Fine-tuning?

Fine-tuning trains the text to image model on specific objects or styles to create a customized model.

How to Fine-tune?

The quality of fine-tuning depends directly on your data quality.

Requirements

  • Images: 10-20 images. The minimum amount is 10 and maximum is 20
  • Image Content: For character or object training use only the subject itself, without other characters and objects (Not relevant for style training, but the style itself must be consistent). For style tune try to get wide range of subjects in the same style.
  • Image Format: JPEG or PNG (no transparent backgrounds)
  • SVG export: Avoid gradients/shadows if you plan to export as SVG
  • Resolution: around 1024px is ideal, max is 1328px
  • Consent: Don't use images of people without permission. Do not use copyrighted images as training data.

Sample Image

Use simple images where it is clear what is on the image (example 1). Images with weird poses or abstract meaning are not too good (example 2).

Example 1 (good):

training data good example

Example 2 (bad):

training data bad example

Leave white space around subjects to prevent edge cutting

image format example

10 consistent, well-formatted images are better than 20 mediocre ones

The training process takes about 1.5 - 2.5 hours (depending on the number of training images).

Using Your Model

After training completes, the model appears in the styles dropdown on thegenerate page. Sometimes it may be necessary to guide the model further using prompts to achieve the desired style. if the training data consisted of black and white images with pink accent color but the generations are not following this style then you can help the model by describing the style and colors in the prompt. Example: a man riding a bike, black outline, pink clothes, doodle illustration

Custom-trained models run slower than our premade ones, with image generation taking around 20-30 seconds.

In the image editor, the Edit tool works best with pre-made models. Fine-tuned models may have mixed results.

If there are any complications or if you need help, feel free to contact me at kristjan@ilus.ai or at my Twitter/X

Start Fine-tuning