Training on 20k images

Training on 360k images (Entire MIMIC dataset)

Inversion of images not seen during training (old model)

Real

Inverted

Brains - works really well (qualitatively)

• Trained on 8,000 brain slices from 5 datasets
• Problem is easier though because brains are already registered rigidly, so it removes a lot of variation due to pose, …)
• Also resolution is smaller, 256x256

Real

Fake

Microscopy images - also works quite well (qualitatively)

• PANDAS 2020 dataset with prostate cancer (11k images, augmented through rotations to 70k)
• 512x512 model resolution, original files have larger resolutions, but variable across images (e.g. 1500x1700, 1650x1900, 400x1800, etc …).

Fake

Real

Fake

Real