#2 Training a diffusion model with only serif letters

In experiment #2, I wanted to find out whether individual font categories could be generated by the diffusion model. This is what came out:

The model was trained with almost 1,400 A‘s in serif fonts. The result is similar to the previous experiment: at 20,000 and 30,000 training steps, some samples are not yet recognisable as A‘s. From 50,000 training steps, the model reliably generates more and more new A‘s with serifs, and at 250,000 training steps, there are almost perfect black A‘s on a white background. The output in this experiment is very similar.