To train a custom an AI that draws like me I used CycleGAN. I trained the GAN with 28 sketches of mine over top of flemish plein-air and early impressionist paintings that were dissected into unique 256 crops. These 28 paintings were obtained through the MET Open Access API. I drew over top of them using the Concepts app on iPad. Here are some examples of A and B inputs:
“The Beeches”
“A Bird's Eye View”
“Boating”
“The Road from Versailles to Louveciennes”
I trained these 256 images for 400 epochs with 256 inner loops each.
Sample of various inputs (photographs and stable diffusion outputs) run through my GAN drawing AI:
There a 6 samples here arranged in a grid where each sample is: [ # input | # AI drawing prediction | # painting prediction ]
So the entire grid is:
| 1 input | 1 AI drawing | 1 AI painting | 2 input | 2 AI drawing | 2 AI painting | | --- | --- | | 3 input | 3 AI drawing | 3 AI painting | 4 input | 4 AI drawing | 4 AI painting | | 5 input | 5 AI drawing | 5 AI painting | 6 input | 6 AI drawing | 6 AI painting |
After my AI creates the Cole drawing I run the output through a few cleaning processing filters in Photoshop and Illustrator to generate cleaned up SVG: