Finally got some interesting results out of the [code example](https://github.com/msracver/Deep-Image-Analogy) Microsoft research released for their [Visual Attribute Transfer through Deep Image Analogy](https://arxiv.org/abs/1705.01088) paper.

Don't have time to go into details but I'll do a writeup on this later, for now I just wanted to share some results.

![IMG_3265-style4_r_1.w_2AB.png](https://steemitimages.com/DQmVbpdKFndKaAsEroiVUYGHxsrtFUB3LvH3Sn26vweQMjk/IMG_3265-style4_r_1.w_2AB.png)

![IMG_3265.jpg](https://steemitimages.com/DQmfMNCGEQYxv4hmatMRzKDeAaEgpLMYj4idqTt3QeiB5Cz/IMG_3265.jpg)

Style image
![style4.png](https://steemitimages.com/DQmSqjYpSHXw4XSAvMArW23t5pf4YtoNxKHrgNwSRbg23Pk/style4.png)
["miss universe" by Silvia Pelissero](http://agnes-cecile.deviantart.com/art/miss-universe-353517674)

Reverse style
![IMG_3265-style4_r_1.w_2BA.png](https://steemitimages.com/DQmbLUsjtfsHrctV9aiotCUZvfWmRbn7EEF9wz9VAYaNw6T/IMG_3265-style4_r_1.w_2BA.png)

Result with the old algorithm (using [neural-style-tf](https://github.com/cysmith/neural-style-tf)) for the same inputs
![result.png](https://steemitimages.com/DQmZjNj1fMM12GSu89Z2mKtLpUPXaXs1FP8KQ6yJsLgP8Rx/result.png)

Turns out to be really hard to get the output you're looking for, have to get the parameters and style/content image combo exactly right. Here's some of the more interesting failure cases:

![IMG_3265-style1_r_1.w_2AB.png](https://steemitimages.com/DQmUB8t9kc1FDXdFe8oes6CmWGadYPCkgsdLJ5f9NVAwBt8/IMG_3265-style1_r_1.w_2AB.png)

![IMG_3265-style6_r_1.w_2AB.png](https://steemitimages.com/DQmanzr15QsSmoevR1DyR7QsZNXV93SZoEJrAHgpeG7RAmC/IMG_3265-style6_r_1.w_2AB.png)

![IMG_3265-style9_r_1.w_2AB.png](https://steemitimages.com/DQmcVXNobZ9bQXGbf2RJTHBVqm1dxD6vCEMYG4aspd3cMQK/IMG_3265-style9_r_1.w_2AB.png)

![IMG_3265-style10_r_1.w_2AB.png](https://steemitimages.com/DQmaCDGHx3iuvPrynKx3g8DBmbhyLpw4uSJTVyNCHdebQR8/IMG_3265-style10_r_1.w_2AB.png)

---

Have a nice weekend!