The future will be full of AI-generated images and news articles. But what about videos? The Meta AI research group just revealed its Make-A-Video tool, which creates short clips based on a written prompt. It can also turn still images into videos, or create variations on existing videos.
Early videos produced by Make-A-Video are pretty rudimentary. They’re smudgy, creepy, and clearly “fake.” This AI also has a disturbing habit—it likes to put human hands on non-human subjects. One video, which was made from the prompt “a teddy bear painting a portrait,” looks like it belongs in The Shining.
According to Meta AI’s research paper, the Make-A-Video AI is trained on public and unlabeled videos. It’s certainly the fastest method of training an AI model, but of course, it means that the AI is somewhat unsupervised. It may make incorrect inferences about a video’s contents, and it may be exposed to racial bias, pornography, and other unwanted material.
If this Make-A-Video ever opens to the public, Meta AI will probably curb these problems by preventing certain words in user prompts.
Additionally, Meta AI faces several technical hurdles with this AI. It needs to produce individual frames of a video that animate together smoothly, which is quite a difficult task. As a result, Make-A-Video can only produce 16-frame clips at a resolution of 64 by 64 pixels. (A separate AI increases the video resolution to 768 by 768, which explains why this stuff is so smudgy.)
The prospects of a video-generating AI are endless. Regular people and professionals could use such a tool to quickly create content. Of course, tools like Make-A-Video may be used to spread misinformation in the future. And at the time of writing, nobody really knows how to avoid that problem.
If you’re interested in Make-A-Video, visit the project’s website to sign up for early access. We’re not sure when the AI will open to the public.
Source: Meta AI