Where’s Sora?

More than half a year since it was first introduced, OpenAI Sora specializing in video creation has not yet been officially launched. This is contrary to the tradition of artificial intelligence start-ups.

The video of the “balloon head” person was created from Sora. Photo: Shy Kids.

Since the appearance of Sora earlier this year, many AI products to create videos have also been introduced. Luma released Dream Machine with a cinematic-grade promotional video. Meanwhile, Runway also gives users a free trial. Unofficial statistics show that China alone has at least 20 start-ups developing video AI.

However, after more than half a year, OpenAI Sora, the most impressive product, has not yet been launched. This is in contrast to the working style of this company. They introduced Chat GPT-4o during a live stream, and users can use it immediately after the event.

Sora is more complex than ChatGPT or Midjourney

There are many AI models for video generation that have been introduced recently, but the number that can be used is very limited. Sora is still in the process of closed testing, only for selected groups of artists and filmmakers. Alibaba and Baidu’s AI Video is only used internally. Kuaishou Kling is gradually gaining popularity, but users also have to sign up for a trial queue.

Current video creation solutions include entering requirements, choosing scale, style, and time. The main technical difficulty lies in the resolution and length of the finished clip. This is also the focus of competition between companies in this field. To have an advantage over competitors, start-ups must invest in strong hardware to increase processing capacity.

The videos introducing Sora used to make a strong impression. Photo: OpenAI.

AI researcher Cyrus said that the models currently available only have a resolution of 480/720p. In addition to the processing capacity of artificial intelligence hardware, training models is also very complicated. When a model that is trained to output low-resolution video is pressed to a high level, it will create many physical and logical errors.

According to Bloomberg, creating videos with AI takes more computational power than running chatbots or making still images. These technical challenges can affect business problems, when customers have to pay too much cost.

Chinese video creation programs currently only support short videos of 2-3 seconds, a simple form of movement. The highest-end version is also only 10-12 seconds. The most high-end product ever introduced is the Sora with 60 seconds, but it’s still in the testing round.

In addition to length, the ability to create story scenes is also the limit of video AIs. “Technically, it is possible to create an hour-long video. But we don’t want scenes like from surveillance cameras, with repeated, motionless scenes,” Zhang Heng, head of research at Pomegranate AI, told Fixed Focus.

Reasons for the delay

At the recently held World Artificial Intelligence conference, AI experts said that video generation is still in its infancy, similar to the GPT-3 era. Therefore, they estimate that it will take at least half a year for it to reach the completion stage. In addition, the complexity of this model compared to the language, also makes it difficult to commercialize on a large scale.

Sora has a longer development time than Midjourney and ChatGPT because it has a timeline element.

In addition, the data source for training the AI video generator is also limited. “We have exhausted our training resources, we are looking for new ways,” Zhang Heng said.

A demo of Kling AI from Kuaishou, a Chinese start-up. Photo: Kuaishou.

Current AI models are only strong in a certain area. For example, Kuaishou Kling works well when users enter commands related to food or broadcasts, because of the large amount of data provided by the parent company.

Sora doesn’t have a public test, but even with the selected OpenAI clips released, it shows that creating objects that accurately comply with the laws of physics is still a big challenge. Responding to Bloomberg, the development team said that they have not set an official launch timeline for Sora.

An executive revealed the time could be the end of the year. Many sources speculate that OpenAI will have to postpone its launch date after the US presidential election. Sora raised concerns about spreading fake videos that the development team has not yet managed.