01: The era of AI abstraction
AI videos have developed to the stage of making people's eyes black, as long as you have watched the following video, you must understand what I am talking about.
In just a few seconds, the AI made a change that turned the cold rice into dung
At first glance, a few seconds ago, you may think that Haojing, who became popular at the beginning of 2023 (his real name is "Sun Guoshuai", who once served as Wu Jing's stand-in) is frying cold rice again, but now you must be kneeling and begging for a pair of eyes that have not seen it.
The creator used an unexpected and unreasonable way to pollute the eyes of the people, and Wu Jing had to call the police when he saw it.
Good news, this is AI-generated, bad news, there are quite a few similar works.
If it is said that using AI video technology to bring old photos that have been sealed for a long time to life is an excellent application of humanistic care, then the recent blowout of AI videos on the Internet is to open the Pandora's box in the era of AI abstraction.
If they don't stop domestic AI, they will go against the sky!!
When you open the push of the short video, from the confusion when you find the kitten eating noodles, to the shock when you find the panda eating, the speechlessness when the tiger eats hot pot, and the helplessness when the lion dances, you can't help but think: "Where did this mobile phone daddy ™ do for me?" ”
"The video can't be P's, so it's real"
This is just an appetizer, with the entry of domestic large models such as Kuaishou's "Keling Kling" and byte's "Dreamina", the creative power of domestic AI videos has been completely released.
In the past, you can only take screenshots and add words to make emoji film and television bridges, but now you can achieve whimsical spoofs with your fingers, and you don't have to rely on PS and AE to spell technical strength.
So you can see Erkang wearing sunglasses, eating ramen, answering the phone, and wearing headphones.
Erkang became the "Changeable Star", but the headphones were molded
You can also watch the intimate Rong Mama who gets along happily with crape myrtle, for fear that crape myrtle will starve to death without a bite.
It can be seen that when I ate the watermelon, crape myrtle was finally full
In the past, a screenshot in the classic film and television scene can be changed according to the indicator word.
Whether it is Niu Hulu who eats ice cream in Zhen Huan's biography, or the emperor who nibbles on chicken legs and forgets to scold the poisonous woman, Elder Tang who raises a gun to physically surpass sentient beings, and Monkey King who uses a smartphone to shake people and catch demons, all made netizens feel that they were in cyber illusion.
With the continuous breakthrough of netizens' brains, even the timeline of normal history can be tampered with by AI.
The Korean drama "The Fifth Republic", known as the "Five Studies", originally had a plot that was in line with the official history - Kim Jae-gyu's "Killing the Sun with Insects" (assassinating Park Chung-hee), but it was made by netizens with AI as "Park Chung-hee counter-killed the assassin Minister Kim", which completely reversed the tradition of the classic "Xia Ke Shang".
Takagi Masao (Park Chung-hee's Japanese name) After all, he was an officer, so it's normal for him to use a gun to fight back
With the emergence of more AI abstract works, the creative potential of netizens has been completely unleashed.
AI may be the most powerful productivity tool, but it will never replace humans when it comes to conceiving evil content.
What's more, it has now passed the stage of throwing bricks and attracting jade, and has directly realized the "human-to-human" transmission.
Following Haojing's play, all kinds of two-player scenes cannot escape the hands of AI.
Wu Jing, the real card in "Slaying the Wolf", has kissed Donnie Yen under the prodding of AI
We still don't know what happened to Hua Qiang and the melon stall owner that day.
It can't explain whether the relationship between Hua Wujian and Xiao Yuer is friendship or love. ‘
Not surprisingly, the picture of Li Yunlong kissing Chu Yunfei will be seen soon, and maybe in a few days, various spoof videos of same-sex/opposite-sex friends kissing affectionately will appear in the WeChat group chat.
There is a B station up master "Mike" 1:1 restored CCTV's seven sets of classic programs "Get Rich", providing new ideas for the future development direction of rural revitalization in the mainland......
Look at these vivid dinosaurs in front of you, not only look cute, but also have high economic benefits, through the greenhouse breeding dinosaurs, many farmers have gone to the road of poverty alleviation and prosperity......
Compared with dinosaur breeding, the domestication of aliens is more challenging, but with the continuous efforts of mainland farmers (only a few dozen villagers have been lost), the feat of domesticating aliens has finally been achieved.
With the blessing of AI lenses, the originally ferocious aliens and dinosaurs have suddenly become approachable, which has greatly improved the acceptance of these new species by the majority of farmers, paved the way for them to enter the farm on a large scale in the future, and well combined the cutting-edge AI technology with the practical application of agricultural and animal husbandry production......
If the above AI spoofs can be regarded as abstract play within the normal scope, then content like "Tiangong Fairy Version Ultraman" can make you completely convinced of the creativity of carbon-based creatures, and the impact of Yunnan poisonous mushrooms on the human brain is still too great after all.
"Always smiling"
So behind this carnival, how does the domestic large model exert its strength?
02: Feng Shui takes turns
Since AI technology has entered the public eye, the entire domestic AI circle has long had an embarrassing sense of "hard work without merit".
Even though domestic benchmarking products such as "General Large Language Model" and "Wensheng Graph Large Language Model" have been launched one after another, in public opinion and user perception, domestic AI has been crushed by OpenAI, and many times it has been treated as an "artificial intellectual disability".
The well-known "artificial retardation", A.K.A. Cyberborn (brute) - comments Robert
Especially when "Wenxin Yiyan" was first launched, it was still unstable, relying on an AI picture of "fish-flavored shredded pork" to face off against Midjourney and DALL· E, making its image extremely tragic in the hearts of netizens.
An episode in which AI is more evil than netizens
And the emergence of this wave of video-generated AI has completely reversed all this.
It's the opposite, and now it's the turn of foreign netizens to ask for a big model account for us.
If you don't have a mobile phone number to log in, you can directly ask someone to help generate it
In the past, we couldn't buy ChatGPT members without overseas credit cards, but now it is repeated in the reverse situation of foreigners begging for Chinese mobile phone numbers.
Since June this year, the domestic AI video model represented by Kuaishou Keling "Kling" has been launched one after another, but it did not attract widespread attention from domestic users at first.
On the contrary, after multiple evaluations by overseas technology media and industry insiders, these amazing generated videos were transmitted back to China, which completely set off a wave of queuing tests.
Industry insiders commented: "China has a huge advantage, everyone who understands it"
Although OpenAI amazed the world with Sora's demo clip at the beginning of the year, even the director of CCTV couldn't help but exclaim, "What should we do?" But it is still in closed beta, and only a very small number of artists and industry insiders can be invited to use it.
After all, the understanding and simulation of the real physical world, as well as the depiction of ultra-high-precision picture details, require huge computing power and resources.
Just the details of the skin texture and the reflection of the sunglasses can give a glimpse of the leopard
It is true that the Sora demo film can still be said to be the benchmark for AI-generated videos, but after all, everyone can't use it, and it can't quench their thirst, so that some netizens who can't bear it are wondering if OpenAI is drawing cakes in PPT.
At the same time, existing large models such as Luma and Runway are aesthetically tired after a long time and want to experience new products.
Domestic AI video models such as "Keling" and "Instant Dream" have seized this window period, on the one hand, they have vigorously practiced the efficiency of the model, and on the other hand, they have been positioned close to the people and reduced the restrictions on their use.
The required AI course generated by "Keling" - Will Smith eats pasta
At this stage, their basic functions are free, "Colin" is not limited to the number of uses, each time can generate 10 seconds of video, and can be extended up to 2 minutes of 30 frames of 1080p video, in terms of architecture, Coline uses the same DiT (Diffusion Transformer) architecture as Sora, and replaces the convolutional network-based U-Net in the traditional diffusion model with Transformer.
In other words, it is to use the efficiency advantages of Transformer in processing and generation to find an equilibrium between fitting ability and parameter capacity, and finally improve the overall training efficiency, preempt other opponents, and take the lead in handing in the papers.
The Transformer model not only ensures the quality of image generation, but also has better scalability and computational efficiency.
"Instant Dream" is a little behind due to the limitation of computing power, and there are certain restrictions on the number of uses and generation time, but after all, it is backed by bytes, and it is estimated that it will soon usher in a big upgrade.
Relying on these technological investments, they have greatly lowered the threshold for Chinese to use AI-generated videos (like "Keling" even if it needs to queue up in the internal testing stage, but the openness is already considered a semi-public test), and there is no need to go to the Internet or understand English.
What's more, we worked hard to conceive the video and make it, and we definitely can't watch it ourselves, these two are directly backed by the two short video platforms of Kuaishou and Douyin, and the AI video generation is completely productized, and it opens up the closed loop of creativity = > production = > sharing = > communication in one step.
Fast forward to my grandfather in the group and ask you: "Why does this Li Yunlong also use a smart phone?" ”
This trend has now spread to station B and video account, maybe it won't be long before AI videos will appear in the chat of a family of people who love each other, and that is the high-risk period for elders to prevent fraud.
"Wensheng video", "picture video", and "video continuation" are the three major functions, of which the high-frequency use is the function of "image generation video", because it is backed by short video applications, the entertainment scene is wider, compared with the generation based on pure text, it is more coherent and the picture is more stable.
Most of the abstract AI videos mentioned above are products of the Tusheng video function, and some of them are presented in the form of Tusheng videos and then spliced into the original video.
Therefore, this creative blowout is not accidental, because Douyin Kuaishou has insight into OpenAI's shortcomings in commercialization (although Sam Altman has no shortage of knife music in the short term, but training large models is really burning money), so that domestic large models can be opened to the natural traffic pool of short video platforms earlier, which truly meets the needs of ordinary people to share their creativity, in addition to earning network popularity, it can also feed back more data to AI.
In addition, the actual effect of this domestic large model is not bad, although it is not as good as the Sora, which is called "industrial grade", but it can roughly be on par with opponents such as Luma and Runway.
Compared with the text model, the public is also much more tolerant of the AI video model, after all, the generation result of the two eyes and one black can be regarded as an unexpected joke, so there is basically no overwhelming "artificial intellectual disability" criticism in this round.
The elderly, the subway, eating mobile phones
If Douyin has relied on the bean bag AI assistant to test the waters, as well as pan-AI functions such as Miyazaki Hayao movie filters and Detective Conan filters, it has gradually allowed users to embrace new technologies and set off many rounds of Internet topics, but this time I really didn't expect that Kuaishou, who was regarded as a "turtle", took the lead.
Because this wave is still Kuaishou's good at "rural areas surrounding cities", and in the global "100 model war", we really need our own AI video model.
03: More than abstraction
In the era of AI, everything changes in an instant, the back wave, the front wave, it is difficult to distinguish between you and me.
In June, when Keling was in the limelight, Luma released the latest Wensheng video model, Dream Machine, which is free for all users to use, which not only adds richer aesthetic style options, but also quickly generates 5-second cinematic visuals. Runway also released the latest Gen-3, confidently claiming that it is one step closer to the "world model".
The official demo screen of Runway Gen-3
Under such fierce competition, the limelight of "Kling" overseas has passed, but fortunately, with the increase of domestic users, more works and topics will surely appear, and everything has just begun.
However, it is still difficult to say that it is easy, compared with the text model, the AI video model needs to consume more computing resources, and the dimension requirements for the model are also higher, so in terms of technology, domestic large models still need to catch up in the short bonus period of productization and differential competition.
This is not only about the competition of technology, but also culturally needs to have an AI video model suitable for Chinese, at least with a sufficient amount of Chinese Internet data to feed, not to be rudely polluted by various data around the world.
Luma, Runway and other models in the Chinese and even Asian materials are either too little, or too stereotypical, many times the Chinese moving and moving the race has changed, either to become European and American faces, or to become South American, after all, the data fed by foreign large models is too mixed.
Foreign evaluators have commented that compared with foreign AI video models, Chinese model developers have a better understanding of local culture, and the content generated by large models can better meet the needs of local users.
Just like many users themselves have found that in the video generated by Keling, the action of the characters eating is very smooth and stable, I don't know how many video data of eating and broadcasting Lao Tie is used behind Kuaishou, and it depends on the Chinese to talk about eating.
Therefore, we really need the existence of a large model based on the cultural background of our country, which can not only improve the accurate response of the model to the needs of domestic users, but also promote the development of Chinese AI art creation.
In recent years, a large number of popular AI works with Chinese characteristics have emerged, which are based on the characteristics of their own culture and use AI to realize more fantasy scenes, giving people a sense of familiarity that is detached from reality but feels right next to them.
In these works, we can see that in the familiar rural villages, the villagers raised the "aliens" in the classic Western science fiction movies, and deconstructed the "space horror" in the way of "The Book of Prosperity", and the aliens have become "fellow villagers".
"The two brothers are alien (business)" Author: Mick
This bizarre but non-disobedient combination is even reflected in the unique humor of the villagers using alien strong acid saliva to neutralize the saline-alkali land.
There are also countless creations that develop in the direction of more spiritual and mysterious, borrowing from the "Bai Ze" in the "Classic of Mountains and Seas" to create a sense of Chinese silence that can only be understood and cannot be expressed, but the "strange" reveals tranquility. (Although the AI rudely recognized it as a white horse)
【New Chinese Style Zhi Monster】Backroom Bai Ze Author: Haiyuan Ye II
The "New Chinese Dream Core" created by drawing on the elements of the "Strange Core Dream Core" creates a yellowed sense of familiarity in the memory, which seems to be a supernatural picture, but it is a life memory element that only Chinese people will resonate with, and finally form a "warm sense of alienation".
【New Chinese Dream Core】Fish Burial - The Customs of the Hometown Author: Haiyuan Ye II
Not to mention the many netizen works who take out old photos at home for AI processing and relive the warm moments of the past.
Whether it is an AI abstract video or a whimsical idea with AI as a pen, it shows the output ability of Chinese people in AI art creation, at least China's unique image creation, and finally it is no longer a visual demonstration film on the official website of the large model.
For a while, the technology can't catch up with TOP1, but at least the technology is moved off the stage and let everyone use it well, AI is used as a tool after all, not for watching other people's wonderful kaleidoscopes.
While there is still opposition to AI-generated content in the past, as AI video becomes more popular, people will use this tool more and more frequently, whether it is to improve productivity or to have fun.
Under the tide of the times, developers and creators need to constantly adjust their positioning and roles, and this round of blowout can be regarded as allowing Chinese people to better embrace and use AI technology.
Whether androids can dream of electronic sheep is unknown at the moment, but the AI monkey brother has already put down the golden hoop stick and used the machine gun.