michael_nielsen

What does it mean that I both love and hate this. Also left wondering about training sets. If you just dump in a bunch of art from every culture/era it seems like you’d jut end up with the equivalent of melted crayons rather than true expression

In a way, that is what happens, too, Or with a transformative network like CycleGAN, train with too mixed a dataset, you might get only recolorisation. In my experience. Like in every art, one needs to work hard to do something special. Everyone has her own way.

In other words, there is lots of technology out there. You can choose what you need and build your own work process. For training, when needed, one can select what best works for you, including using your own photos, paintings, drawings, doodles, generated or processed images.

Models/architectures are equally important as the dataset used for training. Many of the most innovative works in the AIArt field are based on architectures and data built by artists/scientists.

Creativity in ai-art is also about imagining and building a new collection of examples for training the AI model.

Not sure if this has already been done but it might be interesting to recursively train the AI with AI art. Would you end up with something truly original, or just stuff no one would consider art?

There is nothing truly original. Whether you pipe information through a human brain or a neural network, the output is always a transformation of the input based on what the model has learned. An yes, some of us are recursively training models on the output of other models.

To think is to put together random bits of private experience in an orderly fashion. We can learn to do it better.