Generative AI to generate layouts
Trains a Variational Autoencoder (VAE) with datasets produced by CyberAgent/Crello and reconstructs layouts from sampled codes of the latent space.
Click the badge above to run the LayoutVAE in Google Colab.
During preprocessing, layouts related to Image/Text/SVG Shape are extracted from the original datasets, and training/validation datasets are constructed
Green rectangles indicate Images, red ones indicate Text, and gray ones indicate SVG shapes.
After training VAE, we can obtain latent vectors of the training data by encoding and a picture below is 2D graphic by rendering them.
Each color represents labels of format in the training data and we could
observe some groups were constructed because instances of same labels were
located near by the VAE training.
Orange instances represent formats related to documents such as resume, poster and so forth. Light green ones on above area represent a format for zoom background. And blue and purple ones represent marketing banners for SNS such as Instagram, Facebook, TikTok and so forth.
We get samplings from a grid [-2, 2] on the latent space and layouts were generated by reconstructing them with the decoder of VAE.
