A well-funded AI lab with a deep bench of research talent is releasing a powerful new model that generates high-definition video for the film and advertising industries. The company, Moonvalley, on Wednesday launched its first model, named Marey—a nod to early cinema pioneer Étienne-Jules Marey—which could soon help Hollywood studios dramatically speed up production and cut costs.
What sets Marey apart—and has caught the attention of risk-averse studios—is its training data. The model was trained exclusively on video content either owned or fairly licensed by Moonvalley, avoiding the copyright gray zones that make much generative content legally fraught.
Moonvalley is rolling out Marey to a group of 20 trusted filmmakers, some of whom work for major, household-name studios (whose names haven’t been disclosed), according to cofounder and CEO Naeem Talukdar. The version being released is still early-stage, with regular updates and new features planned.
The key differentiator, Talukdar says, is Marey’s native high-definition output—something that’s been notoriously difficult to achieve. “The challenge is that if you want to output in high definition, your inputs have to be in high definition, so you need to be able to train the model on HD footage,” he says. In other words: It demands lots of powerful servers.
By contrast, Talukdar says, most other video generation models are trained on lower-quality 480p or 720p video—and even then, they often compress the data before encoding. That compromises the model’s understanding of fine-grained detail, leading to uncanny or distorted outputs (like misrendered fingers). Marey is designed to overcome that.
It’s also operated differently. Unlike most consumer-facing video generators that start with text prompts, Marey is tailored for professional workflows. Filmmakers can input storyboards or keyframes; actors can film demo reels on their phones, which producers can then use to generate scenes showing the actor from different angles or performing new movements. The model can make subtle edits, like changing the direction of wind in a character’s hair, or adjusting production-quality scenes—filling in background details or background characters (commonly played by extras).
Moonvalley is also running pilot programs with brands looking to train the model on their own imagery and style guidelines. That could allow companies to generate broadcast-quality commercials on demand. “We have a number of smaller boutique brands who suddenly are like, ‘Hey, I can go and make a Super Bowl commercial,’” Talukdar says.
The startup raised a $70 million seed round last year, backed by Bessemer Venture Partners, Khosla Ventures, and General Catalyst. Vinod Khosla is described as an “unofficial” member of the board.
Beyond the top-tier investors, Moonvalley has serious R&D chops. Cofounders Mateusz Malinowski and Mikołaj Bińkowski previously led video research at DeepMind and helped develop the model that later became Google’s Veo 2 video generator. The team includes six other DeepMind alumni, alongside talent from Meta, Microsoft, Google, TikTok, and Snap.
While Talukdar acknowledges that AI tools like Marey could lower production costs, he insists it’s certainly not the beginning of the end for actors—or filmmakers. Budgets, he argues, aren’t likely to shrink; instead, creators will use the savings to pursue bigger, more ambitious ideas. “I think you’re going to see new jobs,” he says. “And more importantly, I think what this is really going to do is empower the creators more than anybody else.”
No comments