💡 [REQUEST] - Neural Style Transfer: change layers to compute style/content

🚀 Describe the improvement or the new tutorial

I tried using code from Neural Style Transfer tutorial to apply a van Gogh style to my photo, and results were way poorer copared to what I've seen on DeepArt.io.

Luckily, I found what's the issue, and how to improve quality.

Original paper states:

Thus, the visually
most appealing images are usually created by matching the
style representation up to high layers in the network, which
is why for all images shown we match the style features
in layers ‘conv1_1’, ‘conv2_1’, ‘conv3_1’, ‘conv4_1’ and
‘conv5_1’ of the network.

and also:

The images shown in Fig 3
were synthesised by matching the content representation
on layer ‘conv4_2’ and the style representation on layers
‘conv1_1’, ‘conv2_1’, ‘conv3_1’, ‘conv4_1’ and ‘conv5_1’

In the tutorial, layer conv_4 is used for content and layers from conv_1 to conv_5 for style.:

# desired depth layers to compute style/content losses :
content_layers_default = ['conv_4']
style_layers_default = ['conv_1', 'conv_2', 'conv_3', 'conv_4', 'conv_5']

At first glance everything looks fine, but indexing patterns are different: in the article (and original VGG model) layers were split into groups between poolings, and in the tutorial a continuous numbering is used.

So layers conv_1, conv_3, conv_5, conv_9 and conv_13 should be used for style and layer conv_10 for content.

That produces way better results.

Suggested change

# desired depth layers to compute style/content losses :
content_layers_default = ['conv_10']
style_layers_default = ['conv_1', 'conv_3', 'conv_5', 'conv_9', 'conv_13']

Existing tutorials on this topic

https://docs.pytorch.org/tutorials/advanced/neural_style_tutorial.html

Additional context

Another suggestion to improve quality: #3731

cc @albanD @jbschlosser

Start with the linked Neural Style Transfer tutorial and its desired depth layer definitions. Compare the tutorial’s continuous layer names with the article’s grouped VGG layers, then verify that the documented style and content layer selections match the cited paper and produce the intended tutorial behavior.