The clumsy beginnings of AI-generated videos

Iartificial intelligence (AI) is already used for various uses in video, in particular to put public figures in inappropriate situations or to make them make false statements – a manipulation that bears the name “deepfake” and turns out, year after year, always easier to achieve. But researchers and companies are working on another project: generating animated sequences ex nihilo, describing the desired result with simple words, such as “a black cat dives into a tub of milk”.

As early as 2022, attempts were successful at Google And Meta. These large companies have chosen not to make these tools public, in particular for fear of manipulation that may result from their use. But others have designed similar tools, such as ModelScope, Kaiber or Runway. With varying success: many of the videos made through them and circulating on social networks are distinguished above all by their imperfections.

Assemblies of plans that are too short and too static, sets that are too flat, approximate details and, above all, awkwardly sketched objects and characters: they recall what Midjourney was capable of, the artificial intelligence that generates images that are now almost photorealistic, it eight months ago. Here and there strange things appear: deformed faces, rolled-up eyes, disarticulated bodies. Internet users complain about it or make fun of it, like this Twitter user who congratulates the author of a very improbable western for having him “officially brain-broken”. Elsewhere, a parody of a beer ad shows a woman, her mouth glued to the metal of a giant can, drinking with the greed of an animal in a burning garden. Another netizen summarizes the tone of the comments: “It’s both the most hilarious and terrifying thing I’ve ever seen. »

Very young tools

By looking carefully, it is possible to unearth convincing videos, often in science fiction universe. A genre that lends itself to it, the AI ​​already managing to represent robots and other strange scenery well. Another video: that of a three-master sailing under a celestial vault with an abundance of details. The deck of the boat is recomposed several times per second, which seems to bring it to life. A tingling characteristic of the Kaiber tool.

In the expert hands of an AI researcher, its competitor Runway Gen-2 has produced an imaginary advertisement extolling the merits of Mars colonization. The plot is skilfully served by the AI ​​plans, happily revisiting the graphic atmosphere of the 1950s, imagining a whimsical architecture.

You have 54.1% of this article left to read. The following is for subscribers only.


source site-30