ИСТИНА |
Войти в систему Регистрация |
|
ИПМех РАН |
||
Style transfer is a problem of rendering one image in the style of another image. Since it involves adjusting style at the expense of content, a natural parameter of this method is the strength of stylization, which is very important in practical applications. Classical optimization based method of Gatys et al. (2016) allows to change this parameter, but is computationally prohibitive. Transformation based methods allow real-time stylization by passing image through the transformation network, but require training new networks for different stylization strengths. We propose a transformation based stylization method, requiring only a single network and allowing to continuously control stylization strength in real-time. Experiments demonstrate that this method gives stylizations of higher quality compared to alternative stylization methods.