This website gathers creative yet unvoluntary visual findings along our two years of research creating ClimateGAN.
These images result from a faulty post-processing, rescaling [0; 1] pixels as if they were [0; 255].
A segmentation model can be creative too you know? Especially if you train it with the wrong loss.
GANs are unstable beasts. They tend to collapse, and God knows where they land when you forget the discriminator.
A random collection of failures and accidents in our quest for depth maps.
There's a dark forest of traps ahead: axing your way through model comparison and multi-modal visual outputs plotting is a journey in itself.
Forwarding a series of noise images through a GauGAN non-trained (just initialized) model. When the center pixel is artificially boosted, the model naturally focuses on that part of the noise, creating this narrower and narrower output circle.