Why do we say Neural style transfer doesn’t learn any parameter? It’s all about training the pixels of an image to make it look artistic. If it doesn’t learn parameter then how will it generate a meaningful image?
Well, whats fundamentally happening here is that you are mixing 2 images so you are not creating it from scratch, right. You are taking outputs from some layers for content and style and then using a loss function, you create an image but this is not the sense of improving a neural network. In that sense it says doesnt learn any parameters.
Right! Another way to state the point is that what is being learned is the actual output image. You don’t modify anything about the actual VGG network: that is just taken as a given function. Notice that each time we run the training, the output is a new image, not a new model. If we then want to “restyle” a new input image, we are starting over from the same place. The previous image we got does not contribute anything other than we might have better knowledge about how many iterations we need to run to get a pleasing result.
1 Like