It seems to use the pix2pix model which is included in the CycleGAN package. I tried the CycleGAN trainer yesterday. It seemed to work, although the output was garbage as I only let one training epoch go by. It defaults to 200 iterations which I’m sure a big GPU could rip thru.