New Video-Generating AI Trained 100 Percent on Public Domain Films

Few tech products have been as broadly contentious as video-generating artificial intelligence.

These complex algorithms, which cleave millions of datapoints together into seconds-long gobs of video, are notoriously trained on proprietary material, leading to widespread ethical and legal concerns. (That's before we even mention how much energy it takes to synthesize an AI video.)

Tech billionaires tend to argue that this is simply the way things need to be — if you want AI, we need to feed it copyrighted books, music, and video.

However, one group of AI researchers is working to prove that argument wrong.

Moonvalley is a Los Angeles-based AI startup offering a "3D-aware" video synthesis model which it claims is 100 percent trained on public domain films.

The startup's flagship product, Marey, debuted in a limited run back in March. It's now out and open to the public, TechCrunch reports, making use of a credit-based system, which is typical of most AI video software.

The company is drawing the attention of big names in the film world, like Ed Ulbrich, a VFX artist and producer who worked on movies like "Titanic," "Benjamin Button," and "Top Gun: Maverick." Moonvalley hired Ulbrich to liaison with film studios back in June, a role he was drawn to by the company's "clean model," as he calls it.

Ulbrich was previously down on generative AI, but says Moonvalley's approach helped change his mind.

"I do think a key thing for me which was the game-changer and why I was just compelled is that at its core the idea of [Moonvalley’s] clean model of an ethically sourced, ethically trained, a legit, proper thing," Ulbrich told Deadline in an interview. "No stolen pixels, no scraping of the internet. It’s done in a great way. And it is so important that it happened."

Similar projects have been launched in other mediums. In June, a team of over two dozen AI researchers trained a large language model (LLM) on openly licensed or public domain data, proving that it doesn't take millions of stolen books to build an AI chatbot.

It was admittedly a ton of work, with the team combing through over eight terabytes of data — roughly equal to about 1,685,461 Bibles — once to format everything, and then again to double-check the copyright status of all the material.

The result was an LLM that more or less stacked up to Meta's Llama 1 and 2 7B, which are admittedly a few years old at this point, but impressive nonetheless.

While it remains to be seen if Moonvalley's data truly was publicly licensed, as they claim, it could offer a strong rebuttal to big tech's narrative of data shortages and necessary plunder.

More on generative AI: "Indie Rock Band" That's Clearly Using AI Claims That "We Never Use AI"