# 6 Training a diffusion model with a more complicated letterform

After testing the diffusion model extensively with the letter A, I wanted to test a more complex letter. The minuscule g is one of the most complicated letters in a font because it is difficult to derive from other letters.

In experiment #6, about 3,500 g‘s were used as training data. Fonts from the sans serif and serif categories were used, as these do not usually differ that much and it had already been shown in the previous experiments that individual font categories can be generated. In addition to the results mentioned earlier, it is particularly noticeable that some g‘s are mirror images. Since this does not come from the training data, the model must have mirrored them itself in order to generate a larger number of training data. This is not optimal for type design, but it can lead to interesting new creations. Another striking feature is this generated image. It was sampled at 20,000 training steps and shows that these images were also generated on the basis of the bedroom model. But this is an isolated case among the generated images.