Hi, I am a DL/ML expert, currently working as a deep learning specialist researcher at a mathematics research center.
I don't have experience with the Theano backend, but if you're using the TensorFlow backend I can help you with this quickly and efficiently, because I use a multi-GPU environment for training daily with Keras+TensorFlow, and I've converted several models so they work with that optimally.
So if you're using the TensorFlow backend please let me know, and based on your resources (what type of / how many GPUs etc.), I can help you find and implement the optimal way of multi-GPU usage for the style transfer script. Likely, all I'll have to do is replicate the computation graph on each GPU and compute the gradients on each in paralell. I can also use the Keras device_lib to detect how many GPUs the machine has, and make sure to utilise all of them, so the script could be used on different machines without modification.
Any questions, please let me know in chat! Thank you