There are practical examples of artificial intelligence, some more viewed, others less, others as interesting as GPT-3 and others as colorful as GauGAN 2. Although the name sounds like a Japanese manga starring robots, the truth is that it really refers to to the latest NVIDIA experiment: an AI that creates realistic drawings from our sketches or a description.
Using it is as simple as access this website from the browser and let your imagination fly. Do you want to paint a desert? Ahead. Do you want to paint a landscape with two trees, mountains in the background and a river? Please don’t be shy. The result is sensational and from Xataka we have been able to verify it for ourselves.
I’m not an artist, but GauGAN 2 is
The point of this article is not to get technical and explain how this AI works on the inside. Let’s stay with the idea that the model behind GauGAN 2 has been trained with 10 million high resolution landscape images using one of the ten most powerful supercomputers in the world: the NVIDIA Selene. It is, in short, very capable.
The tool presents us with a blank canvas that we can fill using different tools: buildings, ground, landscape and plants. Each has its own ramifications. In the same way, we can write (in English) a description so that the AI generates a realistic image from it.
In the screenshot above you can see a grassy floor, a river, a couple of trees, some flowers, some mountains, a sky and some clouds. I know, it looks like a quick drawing done with Paint, but when you tell GauGAN 2 to generate the image the result is that amazing.
Do we make a more coastal image? In this capture you have the sand floor, the sea, a hut with its roof, a fence, some moss, some rocks and the sky with its clouds. The result? The image speaks for itself.
Let’s try something else. Let’s try to make a waterfall in a forest. This is the simplest, since we simply have to draw the mass of trees and the water where we want it to be. The AI perfectly interprets that we are drawing a forest and, in addition, it perfectly integrates the water.
As we can see, the result, although it is far from perfect, is spectacular. But it can go further. If we feel inspired or prefer not to draw, we can uncheck the “segmentation” box and check “text”. That way we can describe what we want it to draw. Here are a few examples.
The images, far from being completely perfect, could easily slip through a photo. There are some glitches, poorly integrated things, and nonsensical objects, but the level of approximation that NVIDIA has achieved with its AI is sensational. Those who want to try the tool can do it for free from the project website.