评论
打印
But when you look closely, what do you see?
但如果仔细观察一下，你看到了什么？
There's no single creature in these images.
这些图片里的都不是单一生物。
And AI augments my creative process by allowing me to distill and recombine textures.
人工智能让我能够提取和重新组合纹理，以此来增强我的创作。
And that's something that would otherwise take me months to draw by hand.
如果手绘的话，这要花上我几个月的时间。
Plus I'm actually terrible at drawing.
再说，其实我画画水平很差。
So you could say, in a way, what I'm doing is a contemporary version of something that humans have already been doing for a long time, even before cameras existed.
可以说在某种程度上，我正在做的是相机出现之前，人类一直在做的事情，但是是现代版本。
In medieval times, people went on expeditions, and when they came back they would share about what they saw to an illustrator.
在中世纪，人们外出探险，回来会把他们所看到的事物告诉画师。
And the illustrator, having never seen what was being described, would end up drawing based on the creatures that they had previously seen and in the process creating hybrid animals of some sort.
画师从来没有见过探险家描述的东西，由此会根据他们之前见过的生物来绘制，再创造一些合成动物。
So an explorer might describe a beaver, but having never seen one, the illustrator might give it the head of a rodent, the body of a dog and a fish-like tail.
因此，探险家可能会描绘一只海狸，但由于从未见过海狸，画师可能会把它画成一只长着啮齿类动物头部、狗的身体和鱼尾巴的动物。
In the series "Artificial Natural History", I took thousands of illustrations from a natural history archives, and I fed them to a neural network to generate new versions of them.
我从《人工自然历史》系列的自然历史档案中提取了数千幅插图，并将它们输入神经网络，生成新版本的插图。
But up until now, all my work was done in 2D.
但到目前为止，我所有的作品都是以2D形式完成的。
And with the help of my studio partner, Feileacan McCormick, we decided to train a neural network on a data set of 3D scanned beetles.
在我的工作室伙伴费利肯·麦考密克的帮助下，我们决定基于3D扫描的甲虫数据集训练一个神经网络。
But I must warn you that our first results were extremely blurry, and they looked like the blobs you see here.
但我必须提醒你，我们得到的第一个结果非常模糊，就是这样的一坨。
And this could be due to many reasons, but one of them being that there aren't really a lot of openly available data sets of 3D insects.
造成这种情况的可能原因有很多，但其中之一公开可用的3D昆虫数据集很少。
And also we were repurposing a neural network that normally gets used to generate images to generate 3D.
我们还改变了神经网络的功能，将这种通常用于生成图片的神经网络用来生成3D图像。
So believe it or not, these are very exciting blobs to us.
信不信由你，这一坨对我们来说都是非常令人兴奋的。
But with time and some very hacky solutions like data augmentation, where we threw in ants and other beetle-like insects to enhance the data set, we ended up getting this, which we've been told they look like grilled chicken.
但随着时间的推移以及一些另辟蹊径的产品的出现，比如数据增强，我们输入了蚂蚁和其他类似甲虫的昆虫的数据来增强数据集，最终得到了这个，人们说它们看起来像烤鸡。
But hungry for more, we pushed our technique, and eventually they ended up looking like this.
但我们还渴望做到更好，并且改进了我们的技术，最终得到了这样的结果。
We use something called 3D style transfer to map textures onto them, and we also trained a natural language model to generate scientific-like names and anatomical descriptions.
我们使用一种名为“3D风格迁移”的技术将纹理映射到输出结果上，我们还训练了一个自然语言模型，来生成科研风的名字和生物结构描述。
And eventually we even found a network architecture that could handle 3D meshes.
最终，我们甚至发现了一种可以处理3D网格的网络体系结构。
So they ended up looking like this.
最后的结果是这样的。
