I chose conv_1234 for style because it felt like it had the best balance of color and texture, and conv_3 for content because it captured the edges and shapes, but wasn't overfitting every detail.
Initializing with different noises was maybe my favorite experiment, and I like how different patterns emerge. For example the eyes on the content comparisons are very different, and the input noises were very different in that patch of the image.
I prefer noise initialized images, maybe I'm just in the mood for texture though.
Grid!! I ended up using lr=0.1, style_weight=500000, content_weight=2. I found that in general the content sets in slower than the style, and so especially for noise-initialized images I bumped up the content_weight.
