Shulou(Shulou.com)11/24 Report--

This article comes from the official account of Wechat: SF Chinese (ID:kexuejiaodian), author: SF

With the popularity of ChatGPT, artificial intelligence seems to have suddenly entered our real life, quickly showing great dominance in various fields. With artificial intelligence, from writing to design, from driving cars to regulating indoor temperature, it seems that humans do not have to do a lot of work. Many people think that the era of artificial intelligence is coming. But when it comes to painting, artificial intelligence always stumbles, especially when it comes to drawing hands.

/ tr. by Alex Hughes (Alex Hughes)

Editor | Zhao Jiaming

You might think that it doesn't take much skill to draw hands. But no matter how researchers improve the performance of the AI image generation tool, there is always a problem with the manpower drawn by AI.

For example, this picture, provided by the friend friend of the editor, is generated by AI drawing platform Midjourney. At first glance, it looks very realistic, as if it was really taken at a carnival party. But if you take a closer look, the stuffing will be revealed! In this picture, many characters have strange hands, such as the lady on the right.

Similarly, the hands "drawn" by the AI image generation tool are similar: in these pictures, some people have three hands, some have arms longer than gibbons, and they have seven fingers on their palms.

Why would AI make such a low-level mistake? In response, Peter Bentley, a computer scientist at University College London, explained that these AI drawing tools are based on two-dimensional models, while they have no knowledge of the three-dimensional appearance of the real world.

The hand looks simple, but there is no advanced technique to draw it. However, the dexterous hands of human beings can be said to be the most complex "machines", and it is not easy for professional painters to draw every detail of their hands. For AI, which does not understand the real three-dimensional structure of the human hand, this is simply impossible. (photo source: Pixabay)

If you try to get AI to generate a simple hand image, that may not be a problem. But if you want a complex image with a human hand and deliberately give detailed drawing requirements, artificial intelligence may be confused. The resulting image may be as elusive as an abstract painting.

AI also can't handle the details, but with the explosion of the ChatGPT network, AI drawing has also attracted a reputation: many people think that AI drawing is fast and good. But the reality is that there are "fatal" flaws in these pictures.

"these pictures generated by AI really can't stand careful observation, and people can always find something that goes against common sense." Bentley explained, "this is because, during the training process, the researchers reused a large number of real pictures to train AI, but did not tell them the true logical relationship between these pictures. So for the AI image generation model, all it can do is present what it thinks is the best picture combination."

This is a picture generated by Midjourney, it looks beautiful, doesn't it, but if you look closely, you can see the "hard wound" hidden in the picture. Did you find out? You can leave us a message.

The future of AI drawing, in fact, the artificial intelligence image generation model is gradually improving, but there is still a long way to go to generate a truly satisfactory image. So, how to improve the performance of artificial intelligence image generation model? What is the future of artificial intelligence drawing technology?

Bentley gives us some confidence. "through 3D model training, developers are gradually making the image generation model understand the true structure and logical relationship of the image," Bentley said. in the future, even if we give very complex image description instructions, the resulting image will still be clear and accurate. "

However, it takes a lot of time to obtain enough 3D image data. As we all know, most of the images on the Internet are two-dimensional planar images. Now, most of the images that R & D personnel provide for the artificial intelligence drawing platform are these images which lack detailed information. "

Fortunately, the American company Open AI, which developed ChatGPT, has begun to try to use Point-E technology to solve this problem. Using Point-E technology, a 3D model can be generated by inputting text prompts. At present, this technology has been open to the public, but can not be happy too early, from the current test results, Point-E technology needs to be further improved.

Once the Point-E technology is more sophisticated, not to mention generating three-dimensional images, even a new digital world may be established. Although there are many outrageous flaws in AI drawing, we believe that artificial intelligence drawing technology will continue to improve, and the future generation of realistic three-dimensional images or even a completely digital world may no longer be a dream.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.