Abstract
尽管最近用GANs的face image generation很成功,conditional hair editing(什么是有条件的头发编辑)仍是个挑战性问题,由于the under-explored complexity of its geometry and appearance. (human body 还有统一的模式,但是不同发型的话,就压根没有统一范式了。什么短发,长发,卷发,大波浪,小波浪). 在本文中,我们提出多输入conditioned Hair Image GAN, 是一种新型的conditional image generation method for 可交互的肖像hair manipulation. 为了让用户可以对每个影响hair的主要视觉因素major visual factor进行控制,然后作者把hair解开into四个orthogonal(正交的)属性,包括Shape, structure, appearance, and background. (这里面shape和structure什么区别?) 对于四者中的任意一个,作者设计一个对应的condition module来represent, process, and convert user inputs,(用户怎么输入的,以什么形式输入的) and modulate (调制调解) 图像生成的pipeline in ways that
repect the natures of different visual attributes (遵从不同视觉属性的本质,什么本质.). 所有这些condition modules are integrated with the backbone generator to 形成端对端的网络,这个网络允许fully-conditional hair generation from multiple user inputs. 基于此,通过projecting 直观的或者high-level user inputs (例如:painted masks彩绘面具, guiding strokes 引导笔画,reference photos 参考图像)到一个well-defined condition representation建立了可交互的肖像头发editing system that enables 直接的操作hair. 实验结果表明他们的方法在result quality和user controllability上都有superiority.