Why is the AI expansion of the fire called "outrageous" by netizens?

　　Zhongxin Jingwei, December 10 (Lin Wansi) Recently, AI expansion has become popular, with the Douyin topic "AI expansion" accumulating 920 million views, and the topic of "AI expansion is very good, don't expand next time" has been played more than 770 million times on Douyin.

　　Many netizens also shared their expanded photos on social platforms, some said "saved my waste film", and some called "don't be too outrageous", "don't care about people's lives or deaths", "AI is more crazy than people". At present, the market popularity of AI expansion is still fermenting, attracting Internet giants and some listed companies to layout.

What are the reasons for the frequent accidents?

　　AI enlargement is to expand a photo according to the same scale or free angle, and AI will predict and supplement the expanded part according to the content of the image.

　　At present, the usage of AI expanded image, one is to expand the image of a single image, and the other is to use a single image and keep expanding the image.

　　Comparing the two images before and after the expansion, it can be found that the output effect of some photos that need to expand the background range is more reasonable, but when there are additional elements in the original picture that need to be completed, it is difficult for AI to imagine the corresponding images of people and objects according to the details, and the proportion and position are not very accurate.

　　In some of the works uploaded by netizens, the AI extended picture directly modifies the species: for what the person who does not show his face in the photo, the AI extended picture helps you reveal: "The person looks like a dog"; A woman wearing a khaki jacket takes a photo, and the AI extended picture directly "grafts" the upper body to the wooden fence of the same color. There is also indescribable content after the photo AI expands the picture, which makes netizens call "ruining the three views".

　　Netizens upload a half-length photo of themselves, and AI directly turns them into birdmen. Image source: Social media

　　This "face card" looks weird. Image source: Social media

　　A father kisses his mother's pregnant belly, and the AI expanded image uses the pregnant belly as a face. Image source: Social media

　　AI: I know how to make a wish. Image source: Social media

　　Another way to use it is to use a single image and keep expanding. According to the experience shared by netizens on social platforms, in the end, it will become a train, an airplane, a corner of the city and a forest.

　　Some classic scenes of film and television have also been "murdered", such as Zhen Huan's serious expression in the biography, Zhen Huan who is on the way back to the palace, wearing sportswear and sneakers to play basketball in the deep palace after expanding the picture;

　　In fact, the popular AI expansion is one of the many hot tracks in the field of AIGC (generative artificial intelligence). In response to the unexpected situation after the expansion of the image, Lin Huijie, CTO of silicon-based intelligence, pointed out in an interview with Zhongxin Jingwei that this is related to the fact that the AIGC algorithm is still not mature and cannot accurately control the generated image content.

　　Lin Huijie believes that AI extended images are the same as AI Wensheng diagram technology, AI Wensheng diagram needs to generate pictures through the input of semantic information, and the text description itself is relatively generalized, which makes it difficult to accurately control the content presented by AI, and the content understood by AI is also difficult to accurately present human thoughts, unlike human drawing. AI extended images expand some image information on the basis of the original images, which is essentially similar to the technical principle of AI Wensheng pictures.

A number of manufacturers have laid out AI expansions

　　In fact, AI expansion is not new. According to incomplete statistics, Zhongxin Jingwei currently includes related application products and scenarios such as Midjourney, StabilityAI, Adobe, Meitu App, and Wink, which have tested the waters and set foot in the field of AI image expansion.

　　In March this year, Adobe released Firefly, an AI creative generation tool, and announced its full opening at the end of May. This includes the ability to expand existing photos and images with one click. In July of this year, Midjourney updated the "Pan Expand" feature, which allows you to freely pan images back and forth.

　　In mid-July, the Meitu App, a product of Hong Kong-listed Meitu, launched the AI image expansion function. At present, there is still an entrance to the Meitu App homepage for AI enlargement, and each person has 3 free opportunities per day to expand images in different proportions of 110%, 125%, 150%, 200%, and 300%.

　　In addition, the application of AI expanded images on the B-side also includes marketing poster design, e-commerce picture production, game material design, etc.

　　However, at present, there are not many AI expansion applets and websites developed by some individual creators. Zhongxin Jingwei searched in WeChat, and there are not many WeChat official accounts and mini programs named after "XXAI Extended Map", and many of them are not free to use.

Where's the next hot spot?

　　In fact, at the end of 2022, AI painting, another track of AIGC, became a traffic password on social platforms, and was also complained by many netizens that it was "somewhat outrageous".

　　In July this year, Miao Ya camera became popular, users choose more than 20 photos, pay 9.9 yuan, and then choose the template they like, they can generate their own "digital clone", the effect is comparable to the market price of dozens of yuan or even hundreds of yuan of art photos.

　　So far, where is the next hot spot in AIGC?

　　Lin Huijie pointed out that AIGC technology will develop more in the video field in the future. He said that video is the most commonly used, the most accepted by users, and the most imaginative application scenario, which is far greater than the expressiveness, interactivity and imagination of pictures. Compared with images, videos can present content in a high-dimensional way.

　　In fact, AIGC, which exploded this year, also has picture generation videos.

　　Recently, Alibaba launched Animate Anyone, a project developed by the Alibaba Institute of Intelligent Computing, which allows users to animate a static character image and some movements and poses while retaining the detailed features of the character.

　　In mid-November, social media giant Meta released a tool Emu Video, which can generate video clips based on text and image input, and ByteDance released a PixelDance model that can generate videos containing complex scenes and actions through descriptions (plain text) + first frame guidance (images) + end frame guidance (images).

　　In addition, Runway has launched the Motion Brush function in Gen2, which can animate everything that is still by simply swiping at any position on the image, and Stability has launched Stable Video Diffusion, which can generate high-quality video clips from images.

　　In addition, the direction of digital human is also the direction of AIGC technology development and application, and digital human technology can be applied to video scenarios in combination with AI-generated images.

　　In October, the celebrity cross-language translation video went viral on the Internet, in which Guo Degang was interviewed in English and singer Taylor Swift in fluent Chinese, not only the sound was similar to himself, but even the lip shape could be matched, which made many people say that "voice actors are going to lose their jobs." During the Shanghai Film Festival in June this year, silicon-based intelligence successfully "resurrected" the deceased, allowing the late director Xie Jin to meet the audience in the form of digital humans.

(The views in the article are for reference only and do not constitute investment advice, investment is risky, and you need to be cautious when entering the market.) )

All rights reserved, without written authorization, no unit or individual may reprint, excerpt or use in other ways.

Editor in charge: Luo Kun Chang Tao