Artificial Imagination: Deconstructing Style

Preface

Each section shows off experiments with a different model, and as such, trained separately. When listing the epoch number, that is merely to identify stage of training, but all testing images will utilize the model at the latest epoch: 200.

Also, all experiments will be shown in reverse chronological order: that is the most recent ones will be here at the top.

Phase 2: Retraining on Itself

You will notice that the following retrained models only show the training process: this is to show how the model is now affected by reintroducing (and slowly replacing) the dataset over time, so that it generates art that is based on its own art. As such, I am slowly cutting off its imagination, and allowing it to be "inspired" by its own artworks, molded as it continuously weens off of its creator. However, as the creator, I also have the responsibility in shaping its style: I curate the best testing images that make it back into the training database, so I can help guide it to evolved. One could say this is similar to selective breeding, but with a machine's art style instead. 

You'll notice this cool "smoothing" or mellowing out of the artworks produced--I am using the imperfection of the generator as a feature of change over time. Many weird artifacts get smoothed out as I re-feed these works.