GauGAN2 is a unique AI art generator in that it lets you edit your images. But how realistic are the results?
GauGAN2 is a tool for creating AI-generated photorealistic landscapes. Developed by NVIDIA and available as a web demo, GauGAN2 lets you tweak the image as much as you like, using a mix of sketching and text input to guide the AI system.
With a bit of time, you can refine the image by adding or taking away details until it's just right. The images it can produce often do look realistic, but one wrong stroke and the images can start to look like fantasy artworks, rather than real-life landscapes.
Here's what you need to know about NVIDIA's GauGAN2.
GauGAN2 is an AI art generator with a unique editing window. Developed by the research arm of NVIDIA, the same company that makes GPUs, it's available for anyone to use via a web demo.
It's similar in a way to using Dall-E 2, an AI text-to-image generator, where you can type phrases like "sun setting over the desert" and the AI system will create an image based on that description. However, that's not all you can do with GauGAN2; there are several ways you can continue to edit the image.
One option is to use the segmentation map to see an outline of objects in the picture. Using this information, you can change the shape of things like a mountain range by redrawing its boundaries or erasing elements altogether.
You can also use the pencil tool to draw a circle, for example, which might be rendered as a sun. The same goes for drawing a few peaks in the background, which might become a mountain range, especially if you combine it with text input describing the scene.
After giving NVIDIA's GauGAN2 a try, you might want to learn how to use NVIDIA's Canvas app, which is built on the same technology. It's a free app geared towards artists and creatives, however, it only works using one of NVIDIA's RTX GPUs.
Behind the scenes, GauGAN2 was built using generative adversarial networks (GAN). In fact, its name is a play on the French post-impressionist artist Paul Gauguin and GANs.
The reason why it can generate images that convincingly look like real landscapes is that during its training period it was fed 10 million high-quality landscape images. The computing power required to train the model came from the NVIDIA Selene supercomputer, one of the world's most powerful supercomputers.
GANs are finding uses in many applications, including plenty of AI text-to-art generators. The key difference with GauGAN2 is the ability to alter the AI-generated image in a dedicated editor, using tools like pen, fill, and brush to make changes.
The interface is minimal, but not exactly user-friendly, and its retro look might remind you of the early days of Microsoft paint, as opposed to cutting-edge software. If the controls don't seem obvious at first, you can run a quick tutorial or watch a video walkthrough; just click the buttons in the top right corner of the page.
The aim of GauGAN2 is to create photorealistic images of landscapes such as mountains, lakes, and beaches.
There is a library of labels you can choose from when sketching that help further tag the sketch input, so the AI generator can render an image that's appropriate. These include things like fog, hill, stone, snow, and water, to name a few.
It's not hard for things to go wrong, however, and you're just as likely to produce absurd and surreal images. Simply take the pencil tool and draw a few odd shapes, and GauGAN will spit out something that looks more like an experimental piece of art than anything recognizable.
Inputting nonsense phrases in the text box can also render strange outputs. Sometimes, you don't even need to go that far; simply asking for a mountain landscape produced what looks like a weird combination of snow and fur.
In either case, some artists may find a tool like this useful. It might be used to speed up the process of storyboarding or as a quick way to create concept art for a game or video.
While this demo is a long way off from being a polished product ready to use, it does show some interesting potential.
GauGAN2 joins the collection of AI image generators that turn text into images. Unlike other applications, however, you can edit the image in stages, using sketches and text as input, fine-tuning the AI image how you like.
If used right, GauGAN2 can produce fairly convincing photorealistic images of nature, which might be useful for artists who need to populate a storyboard or quickly produce concept art. It's all too easy for things to turn bizarre, however, and a few stray pen strokes can lead to some interesting results.
Garling has a Master's degree in Music and over a decade of experience using creative technologies. In particular, she loves writing about music production, film, and DIY electronics. Outside of writing, you will find her taking photos or editing audio.