After just one year, the AI-generated "eating noodles" has become so natural and smooth? This shocked netizens around the world. 
The generated videos on the right come from the Wensheng video model just launched by Kuaishou. Kling.
It is not a pre-release or a pure demo collection, but a product-level application that is directly open for testing and everyone can apply. Moreover, Keling supports the generation of 1080P videos up to 2 minutes and 30fps, focusing on "one-click conversion" from brainstorming to publishable works. （官网地址：https://kling.kuaishou.com/） 图源：https://x.com/ op7418/status/1799047146089619589
m8Or?refer_flag=1001030103_
The communication group has a maximum of 500 people and will fill up quickly. Now, the screen is full of tql:
Foreign friends who haven’t used it yet can only be anxious and post “please” on social media: 
It is no exaggeration to say that Ke Ling is now “a It’s hard to find a number”:
The news spread to the Silicon Valley venture capital circle, and it triggered a heated discussion.
Stability AI former CEO Emad Mostaque said: "China's AI technology has its own advantages." 463003684918 YC CEO also forwarded the demo generated by Keling on the More vivid and real than Sora:
Prompt: Une personne tapant son meilleur croc dans son hamburger
Video address: https://x.com/AngryTomtweets/status/1799787209651859910
For those who pay attention to AI, they must have seen a lot of Ke Ling in recent days. generated works . This website also clicked into the application channel as soon as possible and obtained the trial qualification. Next, we might as well try it out and analyze the reasons why Keling is so popular. The first product-level application of Vincent Video in ChinaMaybe you still remember this once very popular "Balloon Man" video. Three creators spent nearly two weeks using Sora to create this stunning 1 minute and 21 second short video. However, Patrick Cederberg, who was in charge of post-production, confessed to many problems in the process, such as the color of the balloon changing every time it was generated, some flaws appearing in the footage, etc. 
Sora generates results. Full video address: https://youtu.be/9oryIMNVtto?si=F6oDzvrhzfVcQGeh
For previous video generation models, it is indeed difficult to generate more than 1 minute of content in one go, especially if the screen is required All elements remain consistent.
Fu Sheng, Chairman and CEO of Cheetah Mobile and Chairman of Orion Star, released the "Balloon Man" video he made with Corin and said that it only took "tens of minutes" to create the continuity , a short film with excellent realism and clarity.
Interested readers please click: https://waytoagi.feishu.cn/wiki/GevKwyEt1i4SUVk0q2JcqQFtnRd
The following 2-minute public welfare short film "A Place Far Far Away" is also completely generated by Keling Yes, can you see it?
In the work "Zootopia Racing Competition" by the creator @AIGC Thirteen, these 20 seconds include the generation of fast racing cars (large movements), animal-driven vehicles (conceptual combinations that test imagination), etc. Difficult, but judging from the results, Keling has solved these problems very well: 
Source: Keling creator @AIGC Thirteen
There is another very interesting case. "How to Open the Holidays" created by @八级Mechanics. This 56-second short video took a total of 3 hours to produce and included 23 shots. Then add dubbing on top of Ke Ling's generated results, and the humorous feeling will be there immediately:
After reading these, We should have realized that the influence of the video generation technology represented by Keling goes far beyond simple creation. The implementation of this technology is accelerating in different research fields and industry tracks, providing transformative potential for a variety of tasks from automatic content generation to complex decision-making processes.
Which industries will be changed first?
Traditional game development is often limited by pre-rendered environments and scripted events. Once video generation models are integrated into gaming, the way games are developed, played, and experienced will be innovated, bringing new possibilities for storytelling, interactivity, and immersive experiences. For game developers, one of the most intuitive methods is to generate customized visual effects and even character actions based on user narratives. M In the DEMO below, we can see that users can create an unparalleled body experience with the help of cocoa: source: https: //x.com/dustinhollywood /status/1800056286215553444
Another industry that will be disrupted is film and television production. Traditional filmmaking is an arduous and expensive process that often requires years of effort, extensive equipment, and financial investment. The emergence of video generation technology heralds a new "democratization era" in film production, and the dream of autonomously generating personal film and television works from simple text input is becoming a reality. Now, what we use Keling to generate is a 5-second single-shot clip. As technology continues to evolve, the length of video that users can generate at a time will also increase. For example, in the future we may be able to generate longer video content at once to maintain the coherence and enjoyment of story scenes. The camera techniques may be more advanced, such as continuous long shots. 
The silhouette work below once again proves one point: AI’s understanding and aesthetic level of art, Not inferior to humans at all. Prompt: “A dancer’s silhouette transitions seamlessly through different dance styles, from hip-hop to ballet, in one continuous shot”。 . Picture source: https://x.com/dustinhollywood/status/1799970059957555210
The style of science fiction movies is completely grasped: Source: Keling creator @狗儿李
AI can also inject inspiration into the production of luxury blockbusters: We can take a look at this generated by Ke Ling In the "Honey" commercial, the performance of AI in simulating the close-up of pouring honey is not inferior to that of the professional camera team:
What technologies are behind KeLing? We were unable to obtain enough Sora research and development details from OpenAI’s brief technical report, but the official website of Keling Large Model disclosed more reference information, mainly including data preparation, model architecture, training plan and Several aspects of optimization strategy.
Relying on Kuaishou’s years of accumulation in the field of video technology, the Keling Model team has built a complete labeling system, including basic video quality, aesthetics, naturalness, etc. Dimensions characterize the quality of video data and design a variety of customized label features for each dimension to refine training data or adjust the distribution of training data.
In order to meet the needs of paired video and text descriptions in the process of training Wensheng video model, the Keling Big Model team has self-developed a video description model, which can generate accurate, detailed, and structured video descriptions, significantly improving video generation The model's responsiveness to text commands.
After the high-quality annotation data is prepared, how does the Keling large model obtain the ability to simulate the characteristics and concept combinations of the physical world?
In the overall architecture design, Keling adopts the currently popular Diffusion Transformer (DiT). Traditional diffusion models mainly utilize convolutional U-Net containing downsampling and upsampling blocks as the denoising network backbone. But some studies have shown that the U-Net architecture is not critical for good performance of diffusion models. By adopting a more flexible Transformer architecture, diffusion models can use more training data and larger model parameters. DiT is one of the representative works under this research idea.
In the past few months, a consensus has been formed in the industry that the success of video generation models is ultimately due to the role of Scaling Law. This consensus is based on the findings of the DiT paper. Using Transformer can steadily expand the model size: as the amount of training calculation increases (the training time of the model is extended or the model is increased, or both), the performance will also Increase accordingly.
This means that for video generation models, as long as more computing power and more data are used to scale up, the generation quality will continue to improve.
The reason why Keling can transform users' text prompts into specific pictures, including fictional scenes that will not appear in the real world, is based on a deep understanding of text-video semantics and the powerful capabilities of the Diffusion Transformer architecture . Driven by the powerful modeling capabilities inspired by its self-developed architecture and Scaling Law, Caling can well simulate the physical characteristics of the real world and generate videos that comply with physical laws.At the same time, based on the team’s self-developed 3D VAE network, the large-scale model can generate 1080p resolution movie-level video, whether it is a vast and majestic scene or a delicate close-up, it can be vivid Present. In natural scenes, the light changes smoothly. Tester: @shanshan
Of course, for the video generation model, another factor that must be considered is: video is a kind of visual content with a time dimension, and incoherent content will make users distracted The experience is greatly compromised.
In order to ensure that the presentation of motion in the picture is more reasonable, the Keling large model adopts a 3D spatio-temporal joint attention mechanism to better model complex spatio-temporal motion, and can generate video content with larger motions while meeting the requirements. pattern of motion. Training and optimization strategy
If you have tested it yourself, you will find that Keling supports outputting multiple video aspect ratios for the same content during the inference process. This is because Keling adopts a variable resolution training strategy to meet the needs of using video materials in richer scenes.
At the same time, thanks to efficient training infrastructure, extreme inference optimization and scalable infrastructure, the Keling model can generate videos up to 2 minutes long with a frame rate of 30fps. Video generation is no longer a game of "catching up with OpenAI"2024 is called the year of the explosion of video generation technology, but before Keling, we never saw it Sora-level products are available, and it is unknown when Sora will be available.
In a sense, Keling is the first real "Chinese version of Sora" and brings this technology to a stage where it is usable, easy to use, and practical.
As Fu Sheng said: "This may be the best Wensheng video product you can use in the world today." Anyone who has tried Keling in person will understand that this is by no means an exaggeration. .
Fu Sheng’s video also gave another point of view: "In turn, it also shows that Sora is not a technical breakthrough, but a product breakthrough."
I still remember a few months ago , Sora has raised the technical level of the entire video generation track with its advantages of 60 seconds of continuous video, high-definition picture texture, continuous camera movement, and movement methods, setting off a wave of competition in the Vincentian video track.
We originally thought that the field of video generation would evolve into a technological catch-up between domestic and overseas countries, just like last year’s text model. However, the release of Keling means that the exploration of domestic Wensheng video large model technology has reached a new height, and it has achieved substantial leadership in product implementation. We may not need to play the “catching up with OpenAI” game again.
Some people have judged that China is surpassing the United States in the field of artificial intelligence.
The birth of Keling may mean the beginning of a new era. In the era of generative AI, generating and editing videos may be as easy as using photos on mobile phones today, and the barrier between imagination and reality will be completely broken. 
Due to its popularity, the number of people currently queuing up to test Corin has exceeded 50,000. If you are interested in how to play AI-generated videos, you may wish to follow the "Keling AI Video Account" first to get more high-quality cases. The above is the detailed content of Kuaishou's "Keling" explodes: a huge shock in overseas AI circles, the Chinese version of Sora is hard to find. For more information, please follow other related articles on the PHP Chinese website!
Statement：The content of this article is voluntarily contributed by netizens, and the copyright belongs to the original author. This site does not assume corresponding legal responsibility. If you find any content suspected of plagiarism or infringement, please contact admin@php.cn
During the internal testing process, we also discovered tutorials and demo documents spontaneously created by a community of professional creators, including hundreds of works that can be generated, and also provided guidance on testing dimensions. 
