At the beginning of this article, Shi Chao first came to show you a picture.
And this one.
And this one.
Of course, there is a reason to show you these landscape photos, because these pictures are completely generated by AI technology and are fake...
In fact, Shichao is already tired of various AI technologies, but this automatic image generation technology can really make people shine.
These images not only look like real photos, but they are created based on your ideas.
As long as you draw two strokes on it, and it is a few of these simple strokes, AI can add "billions of dots" of detail to your extremely simple creations...
It's a big spectrum...
And this black technology general thing, from the hands of Nvidia we are familiar with, this is an AI technology called PoE GAN.
PoE here refers to Product of Experts, the expert product, and GAN means to generate adversarial networks.
Don't you understand? It doesn't matter I don't understand...
But you don't need to know the specific meaning, just know that PoE GAN is the ultimate stitching monster, which can not only identify your soul painting, but also generate pictures based on text input.
Just draw two strokes on it, and it will respond immediately and generate a corresponding picture.
As much as you add, it generates as much as it can, and the score is no different.
No matter how the group of demons dances, AI can calmly cope with it.
You don't have to pay attention to the details inside, the AI will automatically identify different areas, such as the wave shape is the mountain peak, the top of the picture is the sky and so on.
If you don't like the way it is now, you can also change the terrain and weather, and you can give you whatever you want.
You can also type directly, and the AI will analyze it in real time and convert the text into an image.
For example, if the mountain peak is covered with snow, the picture will change from the mountain peak to a mountain covered with snow.
Or hit the beach, the picture will immediately appear corresponding to the scene.
But then you hit the clouds, and the sky changes.
The first time I saw it, I was really surprised, but after watching this kind of demonstration, it is inevitable that I will have some doubts about this technology.
For example, did PoE GAN just stitch together the materials that had already been learned to produce these pictures?
Not only will PoE GAN stitch together large patches of trees, mountains, skies and oceans, but will also fill in the details you draw.
For example, a small piece of the puzzle object drawn in the demonstration will be recognized by the AI and generate a stone in the ocean.
Not only is the recognition successful, but the details in the picture also follow the logic of real life, such as there will be waves around the stone, and the sea water is also white.
This detail, a little bit of a thing.
However, in order to be able to understand this thing more deeply, Shichao still went to the Internet to find the official information and instructions of NVIDIA, and wanted to see the specific explanation.
At present, in the existing image synthesis framework, it is no problem to encounter a single type of demand, but if a variety of different needs are put together, then AI will be confused on the spot...
PoE GANs were born to address this limitation by recognizing different inputs and mixing them together.
For example, here is a very intuitive example, when you only choose stick figures, the picture will generate an image according to the current needs.
At this time, I add different colors and text to separate different areas, and the picture will become a different look according to the needs.
When the detailed text description is also checked, the image will further meet people's demanding requirements on the basis of the image that has been generated according to the requirements.
PoE GAN is indeed amazing, but this technology is not recent, Nvidia in 2019 came up with an AI technology, called GauGAN.
The name sounds a bit weird, but the technology does allow people to generate a picture that is fake and real at a threshold that can't be lower.
Later, GauGAN evolved GauGAN2, and it brought a new way of playing, that is, typing and drawing.
As long as you frantically type the keyboard on the text box, the AI will automatically generate the corresponding picture based on the text.
After that is the previously mentioned suture monster PoE GAN, which evolves fast enough...
In general, Nvidia's new technology is indeed very black technology, casual typing can generate a fake picture, if this had been placed a few years ago, it is estimated that everyone will feel that it is nonsense.
As a result, after the developers worked on it, they really turned this magic-like picture into reality.
In the past, you needed to find a landscape photo, either go to the copyright website to find one, or go out and take one yourself.
Now just let the AI generate it, even in the hot summer, you can easily get a beautiful snowy day.
And this process, basically does not take any effort.
ShiChao estimated that the paintings, games and related practitioners of this technology were happy enough to finally stop tired of design backgrounds...
But there are also photographers in the comments section who have expressed concern that people will not have to go out to create beautiful scenery in the future.
In any case, this technology has indeed opened the eyes of many people, and also made Shichao more look forward to what kind of innovation will occur in the future creative process.
Author: Budweiser Editor: Surface Line
Images, sources:
YouTube - Two Minute Papers
NVIDIA Corporation - Multimodal Conditional Image Synthesis with Product-of-Experts GANs
cnBeta - Nvidia's latest AI technology transforms text into lifelike images