Exploring Latent Spaces

In GAN-based image generation, you can explore the latent spaces between one property or another in an image. This can help to create morphing animations of one image into another.

With StyleGAN-t, a new technique from NVIDIA, images can be generated in 0.1 s per image, which is almost getting to real time video. It also makes it easier to pick the optimal combination of images from a set of text generated images with different properties.

By grouping like images together, StyleGAN-t can also help show which images have similar properties. This might not always be trivial and images generated to simulate cancer growths or certain propteins might see a large benefit in applying this tool.