Neural Transfer Implementation & by Github - Koalaliketree

Style Image

Water Color Portrait

by Andrew Wyeth

Content Image

Orianna

by League of Legend

Optimizing content loss at different layers

We feed in random noise as input image and implement content loss on different layer of VGG-19, as we can see the "lower" the layer is , the closer the image looks like the original pixel image

Conv_4

Relatively Noisier.The image still catch the content of the content image

Conv_4

Conv_3

Obviously closer to content image than "conv_4"

Conv_3

Conv_2

This is weird since it's actually brighter and more ambiguous than Conv_3

Conv_2

Conv_1

Almost the same as the content image

Conv_1

Generate image from two different random noise

We feed in different randomly generated noise, and optimized it w.r.t only the content loss. The output looks quite similar

Noise 1

Noise 2

Optimizing Style loss at different layers

We feed in random noise as input image and implement style loss on different layer of VGG-19

Conv_1 + Conv_2

The loss didn't catch all the feature in the style image

Conv_1 + Conv_2

Conv_1 + Conv_2 + Conv_3

The feature of the style image becomes more clear as we can see nose and eye shape in the generated style image

Conv_1 + Conv_2 + Conv_3

Conv_4 + Conv_5

With only "high-level" layer, the feature is still well-captured, but the overall color is not well-capturedin certain area

Conv_4 + Conv_5

Conv_1~Conv5

Both overall color and detailed feature is well-captured

Conv_1~Conv5

Generate image from two different random noise

We feed in different randomly generated noise, and optimized it w.r.t only the content loss. The output looks quite similar

Noise 1

Noise 2

Fine tune

I fine tune the model with number of epoch and style content weight ratio, I found the best style content weight ratio is 100000:1. The best epoch is 300

style weight 100

We can see the style is not very obvious.

style weight 10000

style weight 10000

We can see the style is not very obvious.

style weight 10000

style weight 100000

The style is more obvious

style weight 100000

style weight 1000000

The style is too strong that some content is not well-captured

style weight 1000000

epoch 300

We can see the style is not very obvious.

epoch 300

epoch 300

epoch 1000

epoch 3000

epoch 3000

Try with different style and content

Noise and content image

Noise output

Content output

Favorite Image