I use an encoder-decoder to generate image by style and content,but I found a weird phenomenon.
I use a image itself for content and style criteria. So it should generate an image like itselt in both content and style.The result have a great color bias.
Besides ,the network is extremely unstable.e.g, it can generate such a green pic in this test epoch but generate another pic behave itself in color red in next epoch,why?
I use vgg16 and the first 3 maxpooling output for style loss.
the size doesn’t matters,don’t care it.