In the early morning of December 10, OpenAI officially released Sora Turbo, a high-end accelerated version of the video model Sora. Compared with the first generation Sora, Sora Turbo generates videos faster.

It is reported that the first-generation Sora launched by OpenAI generated a one-second video in an average of 10 seconds. In the live broadcast demonstration, Sora Turbo generated four 10-second videos at the same time, taking a total time of only 72 seconds.

At the same time, Sora Turbo can generate video from text/image/video at a lower cost.

From now on, Sora Turbo has opened all its functions to members who subscribe to OpenAI Plus and Pro without additional payment, but the usage limits of different membership types vary:

Monthly payment The OpenAI Plus membership of US$20 has a total of 50 video generation quotas in a single month; the Pro membership of US$200 per month can generate videos unlimited times in a single month at slow speed and 500 times in accelerated video generation. If a high-resolution video is generated, it can be used The times will be less.

There are also differences in generating videos. Plus members can generate videos with a maximum resolution of 720p and a single video duration of 5 seconds; Pro members can generate 5 videos at the same time with a resolution of 1080p and a duration of 20 seconds.

OpenAI official website shows that Sora has been supported in 155 countries and regions around the world, excluding mainland China and most of Europe.

As Sora became available for use, netizens poured in crazily and the server was overcrowded. In response, OpenAI founder and CEO Sam Altman posted that user registration has been closed and the video generation speed will slow down for a period of time.

The OpenAI team defines Sora It is a creative tool that allows users to generate the desired video through a text description, a picture or a video.

They mentioned that Sora cannot generate a feature film with one click, but needs continuous optimization. In order to introduce Sora intuitively, OpenAI broadcast the Sora conference live and turned it into an online video tool tutorial.

If the user needs to generate a video, he needs to open the storyboard and display four videos in the storyboard to show the video details from different perspectives.

△Storyboards display different angles of the picture

In the storyboards, the user enters the desired video instructions in the description box, and sets the style, screen ratio, duration, number of storyboards, After adjusting the resolution, a video can be generated.

Currently, Sora can support generation of up to 20 seconds and a resolution of 10For 80p videos, three aspect ratios are available: 16:9 / 1:1 / 9:16.

OpenAI introduced that if the user’s video instructions have fewer words, Sora will fill in more details; the more words, the more it will comply with the user’s instructions.

During the live broadcast, OpenAI entered the command "A yellow-tailed crane standing in the creek" in the description box, put this video clip into the front part of the timeline, and then entered "this" in the new video A crane puts its head into the water and takes out a fish" command, and places it at the back of the same timeline. The two videos are not continuous, so Sora needs to generate a transition video by itself to combine the two videos into a complete video.

The results show that Sora generated clear videos according to the instructions, and generated smooth transition clips between the two video clips, making the video coherent and story-telling. However, there is no obvious fish generated in the video, but there is a splash of water when the crane picks up the fish.

△Two videos are combined into one video

In addition, users can also directly upload a picture or video. Sora can generate text descriptions for subsequent videos based on the content. Users can freely Change instructions for subsequent videos.

For example, after submitting a picture of a lighthouse, Sora will later create a card describing the video that will be generated later. The user can change the instructions and adjust the position on the timeline to decide when the generated video will be played. .

△Upload the lighthouse image, and Sora automatically describes the subsequent video

After the initial video is generated, if the user needs to optimize it, the user can change the object through the remix tool, such as replacing the mammoth with a robot, changing Character expressions, etc. To this end, Sora has also set three intensities: subtle, light, and strong to meet users' different changing needs.

△Replace the mammoth with a mechanized mammoth

If the user is satisfied with some fragments of the generated video, they can use the recut tool to edit the fragments that need to be retained, and then expand the video through instructions. Get a new video.

In addition, Sora also has loop and advanced functions blend. The former allows users to loop the video infinitely, Sora can generate details to connect the video from beginning to end; the latter can integrate two completely different scenes into one.

In February this year, OpenAI launched the first-generation Sora for the first time, which can generate high-definition videos of up to one minute based on prompt words entered by the user. Since then, Sora has begun a 10-month closed beta, open only to specific outsiders such as visual artists, designers, and filmmakers.

Until a few hours before the start of this live broadcast, the latest official video display of Sora was released.Outflows from the network.

During the closed beta period of Sora, similar Chinese products such as Keling AI, Jimeng AI, and Conch AI have been recognized by a group of users overseas.

According to foreign website analysis tool Similarweb, the total number of global visits to Keling AI in November reached 9.4 million times, surpassing the 7.1 million times of similar overseas product runway; in the tweet of the Sora Turbo demonstration video that was leaked in advance Next, many foreign users said that its video effects are similar to similar Chinese products.

Sam Altman once said that the update speed of Sora is not as fast as expected because improving the model requires ensuring security and expanding the computing scale.

It is reported that in order to ensure that model training can proceed, OpenAI has cooperated with semiconductor company Broadcom to develop an artificial intelligence chip to run the model, which may be launched as early as 2026.

Sora’s significance to OpenAI is far more than just a video generation tool. During this live broadcast, Sam Altman emphasized that he hopes that AI can understand and generate videos to change the way people use computers, and will also help OpenAI realize general artificial intelligence (AGI).

However, people have different opinions on this statement. Step Star CEO Jiang Daxin once told Smart Emergence that he understood that OpenAI launched Sora to explore iterative multi-modal generation capabilities, so Step Star also followed a similar path to OpenAI to research general artificial intelligence technology; Meta chief artificial intelligence scientist Yann LeCun said It is believed that simulating the behavior of the world by generating pixels is a waste of resources and doomed to failure.

The application time of general artificial intelligence is already in the OpenAI plan. Sam Altman told the media last week that the first application cases of general artificial intelligence will appear as soon as 2025. People can set a very complex task, and AI will use different tools to complete it.

“The initial impact of general artificial intelligence may be small. Ultimately, its impact will be greater than people think. Just like every major technology appears, a large number of jobs will be replaced.”< /p>

Perhaps with the open use of Sora, its impact on general artificial intelligence will gradually become stronger, and OpenAI will also achieve its ultimate goal.