AI 生成虚假视觉效果的能力还不是主流知识,但一个新网站————提供了快速且有说服力的教育。

该网站是 Uber 的软件工程师 Philip Wang 创建的,并利用芯片设计者 Nvidia去年发布的研究成果,制作了源源不断的假人像。其背后的算法在庞大的真实图像数据集上进行训练,然后使用一种称为生成对抗网络(或 GAN)的神经网络来构建新示例。

“每次刷新网站时,网络都会从头开始生成一张新的面部图像,”王在Facebook 的一篇帖子中写道。他在给Motherboard 的一份声明中补充道:“大多数人不明白未来人工智能在合成图像方面会有多好的表现。”

支持该网站的底层 AI 框架最初是由一位名叫Ian Goodfellow的研究人员发明的。Nvidia 对该算法的看法,名为 StyleGAN,最近已开源,并已被证明非常灵活。虽然这个版本的模型经过训练可以生成人脸,但理论上它可以模仿任何来源。研究人员已经在试验其他目标。包括动漫人物、字体和涂鸦。

StyleGAN 动漫人脸插值是可靠的:

正如我们之前在The Verge讨论过的,像 StyleGAN 这样的算法的威力引发了很多问题。一方面,这项技术有明显的创造性应用。像这样的程序可以创建无穷无尽的虚拟世界,并帮助设计师和插画家。他们已经在引领新类型的艺术作品。

然后是缺点。正如我们在关于 deepfakes(使用 GAN 将人脸粘贴到目标视频上,通常是为了创建未经同意的色情内容)的讨论中看到的那样,大规模操纵和生成逼真图像的能力将对现代社会如何看待证据和信任。这种软件对于创建政治宣传和影响力活动也非常有用。


The ability of AI to generate fake visuals is not yet mainstream knowledge, but a new website — — offers a quick and persuasive education.

The site is the creation of Philip Wang, a software engineer at Uber, and uses research released last year by chip designer Nvidia to create an endless stream of fake portraits. The algorithm behind it is trained on a huge dataset of real images, then uses a type of neural network known as a generative adversarial network (or GAN) to fabricate new examples.

“Each time you refresh the site, the network will generate a new facial image from scratch,” wrote Wang in a Facebook post. He added in a statement to Motherboard: “Most people do not understand how good AIs will be at synthesizing images in the future.”

The underlying AI framework powering the site was originally invented by a researcher named Ian Goodfellow. Nvidia’s take on the algorithm, named StyleGAN, was made open source recently and has proven to be incredibly flexible. Although this version of the model is trained to generate human faces, it can, in theory, mimic any source. Researchers are already experimenting with other targets. including anime characters, fonts, and graffiti.

The StyleGAN anime face interpolations are solid:

— gwern (@gwern) February 12, 2019

As we’ve discussed before at The Verge, the power of algorithms like StyleGAN raise a lot of questions. On the one hand there are obvious creative applications for this technology. Programs like this could create endless virtual worlds, as well as help designers and illustrators. They’re already leading to new types of artwork.

Then there are the downsides. As we’ve seen in discussions about deepfakes (which use GANs to paste people’s faces onto target videos, often in order to create non-consensual pornography), the ability to manipulate and generate realistic imagery at scale is going to have a huge effect on how modern societies think about evidence and trust. Such software could also be extremely useful for creating political propaganda and influence campaigns.

In other words, is just the polite introduction to this new technology. The rude awakening comes later.