Machine learning based art using the Keras interface to TensorFlow.
Over-fits a convolutional neural network to the inverse of a downsampling process on a source image:
The resulting model is then an up-sampler, with the characteristics of the source image imprinted upon it. One can then apply this upsampler to generate art in the style of the source image:
By applying the model as a stencil onto a target image, up-sampling a series of small chunks of the target image and compsiting the result, one can imprint the style of the source image onto the target image:
To train a model on a single image: train.py source.jpg (will produce a model.save)
To train on a set of images: train.py directory/with/images
To predict on some pre-determined and randomly-generated inputs: predict.py model.save
To predict on the above, as well as a particular target: predict.py model.save target.jpg