Tech companies are pushing ahead with ways to make AI-generated videos, whether there’s an audience for it or not.
What happened: Meta announced Movie Gen, its attempt at an AI video generator. Using a text prompt, the model can create videos from scratch, edit existing clips, and generate audio to go with the action.
- Elsewhere, Tim Brooks — who was co-leading development on OpenAI’s Sora video generator — left the company this week to pursue new video projects at Google’s DeepMind AI research division.
Zoom in: At first glance, Meta’s videos look adequate — apart from some fish vanishing while they swim behind a baby hippo, they seem to be free of many giveaways that videos are AI-generated. But what companies choose to show doesn’t always match up with real-world performance, like that Toys ‘R’ Us commercial made with Sora.
-
The clips also suffer from common textures, lighting, and stiff movements/facial expressions that simply “look fake.”
- A woman painting seems to be blasted by bright light in one clip. But while her arm casts a shadow, the easel in front of her does not, and another shadow seems to be cast by nothing.
Why it matters: Video generation is a major money-making proposition for AI companies. It has been pitched as a way to take on tedious post-production tasks, help companies make cheaper ads, and democratize creation for those who can’t afford to hire big crews.
Yes, but: Besides developers, few are stoked on the idea of AI-generated videos. Talent working in film, TV, and commercial production have been wary of the tech, whether that’s because the quality isn’t anything to get excited about, or out of concerns that it will be used to replace talent with a cheap alternative, as seen in recent entertainment industry strikes.
- Many creators, from big studios to YouTubers, have also raised the alarm over generators being trained on their content without permission.
Bottom line: Like other video generators, Movie Gen is only available to a few partners for the moment. So we have to take Meta at its word when it claims its 30 billion parameter model outperforms the competition on video quality — and that it’ll be the kind of thing people will want to use.