Tech companies compete to develop video AI

OpenAI on Sunday unveiled a new text-to-video model, Sora, which can produce up to 1-minute- long videos from text prompts.

Global big tech companies are fiercely competing to develop video artificial intelligence (AI) after OpenAI unveiled its video production AI model, Sora, and are currently strategizing to secure the top spot in the market as the significance of video content continues to grow.

According to The Information on Sunday, ByteDance, which operates TikTok, is developing an AI model that generates 3D avatar videos based on text input. This solution involves advertisers inputting text into a prompt window, with the AI model generating video scripts accordingly and creating 3D avatar videos, thus integrating AI influencers into TikTok advertisements.

TikTok “is now in discussions with advertisers about an AI-powered feature that would generate avatars to star in videos – virtually influencers that would potentially compete with human creators for ad deals,” The Information said. It added that video AI incurs significant inference costs, and the company thus seems to be preparing attractive services for advertisers.

Adobe, famous for Photoshop, is also foraying into video AI development and has now begun acquiring various human facial expression videos from video creators. According to Bloomberg, the cost of purchasing video can range from an average of $2.6 per minute to as high as $7.25.

Adobe, primarily an image tools company, developed its own image AI including Sensei in response to the competition, but faces intensified competition particularly after OpenAI introduced Sora. It is reported that Adobe is expected to launch a new AI by the end of 2024 despite shareholder concerns.

Google launched a video production application called Google Vydra at the Cloud Next 2024 event held earlier in April 2024.

For its part, Vydra is an AI for creating videos for PowerPoint presentations. When a user inputs the sentence, “Make an educational video,” for example, an editable video is generated within seconds. This solution integrates with workspace apps such as Drive, Meet, Docs, Sheets, and Slides, allowing both enterprises and individuals to convert their images and videos into other videos. Google Vydra is currently being tested by some users, with plans to expand its user base.

[Graphics by Song Ji-yoon and Lee Eun-joo]

Video AI has traditionally been the specialized domain of video AI companies such as Runway and Pika Labs, offering services that generate videos ranging from 4 to 15 seconds, but the landscape shifted when OpenAI unveiled Sora in February 2024. Particularly noteworthy was OpenAI’s demonstration of high-quality videos, including a minute-long video of a woman walking in downtown Tokyo and old California Gold Rush video, which differed significantly from previous video AI capabilities.

Google DeepMind chief executive officer Demis Hassabis reportedly told colleagues that it could be difficult for Google to catch up with Sora. Google Research, Google’s research organization, previously introduced its Lumiere video AI model in a paper but hinted that it still had a long way to go.

Google subsequently raised concerns about OpenAI’s data training practices, with YouTube CEO Neal Mohan pointing out that using YouTube videos for AI model training without permission violates the service’s terms of service.

The rush of big tech companies into video AI development is due to the significant role videos play in content, with 80 percent of current online traffic coming from video content. According to Grand View Research, the video AI market is projected to grow at an annual average of 19.7 percent from $472.9 million in 2022 to $1.8 billion in 2030.