# Neural network draws cats
I trained a neural network based on the transformer architecture to draw cats:
- two tokens represent the delta in the XY-axis
- a newline completes a curve
I interrupted the training process three times and generated test images that can be described as learning steps:
1. scribbles
2. shapes
3. objects (cats)
## Scribbles
![[cat-scribble-1.png]]
![[cat-scribble-2.png]]
![[cat-scribble-3.png]]
![[cat-scribble-4.png]]
![[cat-scribble-5.png]]
## Shapes
![[cat-primitive-1.png]]
![[cat-primitive-2.png]]
![[cat-primitive-3.png]]
![[cat-primitive-4.png]]
![[cat-primitive-5.png]]
## Objects (Cats)
![[cat-final-1.png]]
![[cat-final-2.png]]
![[cat-final-3.png]]
![[cat-final-4.png]]
![[cat-final-5.png]]
![[cat-final-6.png]]
![[cat-final-7.png]]