Moonvalley’s marey is a state of the art AI video model that was trained on FULLY licensed data

A few years ago, there was no such thing as a “generative AI video model.”

Today, there are dozens, including many capable of rendering ultra-high-definition, ultra-realistic Hollywood-caliber video in seconds from text prompts or user-uploaded images and existing video clips. If you’ve been reading VentureBeat over the past few months, then you’ve probably come across articles on these models and their companies, from Runway Gen-3, Google Veo 2, OpenAI’s Sora, Luma AI’s Pika, to Chinese upstarts Kling, Hailuo, and more. Even Alibaba as well as a startup named Genmo offer open-source video model.

These models have already been used in major blockbusters from. From HBO’s True Detective Night Country (19459051) to Toys R’ Us’ music videos and TV advertisements, you’ll find it all here: Everything All At Once Coca Cola. Copyright concerns are a major issue that could still arise despite Hollywood and filmmakers’ rapid adoption of AI.

From what we can tell, since most AI video model startups do not publicly share their training data, they are trained using a large number of videos that have been uploaded to the internet or collected from other archives, including those with copies rights whose owners may or might not have given express permission to AI video companies to use them. Runway is one of the companies that are facing a class-action lawsuit (which is still working its way through the court system) over this issue. Nvidia also scraped YouTube videos for this purpose. The debate continues over whether scraping videos and data constitutes fair use.

There’s now a new option for those who are concerned about copyright or don’t want to use models that have a question mark. A startup called Moonvalley – founded by former Google DeepMinders, researchers from Meta, Microsoft, and TikTok among others – has introduced Marey, an AI video model that is generative and designed for Hollywood studios. Marey, a “clean”state-of-the art foundational AI video models, is trained exclusively using owned and licensed data. This offers an ethical alternative to AI model developed with scraped content.

“People claimed it was not technically possible to build a cutting edge AI video model without scraped data,” said Moonvalley’s CEO and cofounder Naeem Telukdar in a video call interview conducted by VentureBeat. “We proved that it was not possible.”

Marey is available on an invitation only waitlist and joins Adobe’s Firefly Video, which the long-established software vendor claims is also enterprise-grade, having been trained on licensed data and Adobe Stock (to the dismay of some contributors). It also provides indemnification to enterprises for using. Moonvalley provides indemnification for Clause 7 of this documentstates that Moonvalley will defend its clients at its own expense. Moonvalley hopes that these features will make Marey attractive to big studios – even as other companies such as Runway have made deals with them – and filmmakers among the ever-growing list of AI video creation options.

More ‘ethical AI video’?

Marey was created as a result of a partnership between Moonvalley, based in California, and Asteria – a studio that specializes in AI film and animation. The model was designed to provide filmmakers with new tools and AI-driven video production standards while maintaining industry standards.

Talukdar stated, “Our conviction was you won’t get mainstream adoption in the industry unless you work with them.” “The industry has made it clear that we need to create a clean model in order to get them to use these models,” Talukdar said. Up until today, it was widely believed that you couldn’t.

Instead of scouring the internet to find content, Moonvalley developed direct relationships with creators in order to license their footage. The company spent several months establishing these partnerships to ensure that all data used for the training was legally licensed and acquired. Moonvalley’s licensing policy is designed to compensate content creators for their contributions.

Talukdar stated that “Most of the relationships we have are coming in now that people know about what we do.” “For small-town filmmakers, much of their footage just sits around. We want to help them monetize their footage, and we’re interested in artist-focused models. It’s a good relationship.”

Talukdar said to VentureBeat that the company is still assessing its compensation models and revising them, but it compensates creators on the basis of the duration of the footage they provide, paying an hourly or a minutely rate as part fixed-term licensing contracts (e.g. 12 or 4 months). This allows for recurring payments to be made if the content is used again.

This company’s goal to make high-end production video more accessible and cost effective, allowing advertisers, studios, and filmmakers to explore AI-generated story telling without legal or ethical issues.

More cinematographic control – beyond text prompts and images

Talukdar explained Moonvalley’s Marey AI model took a different approach than existing AI models by focusing more on professional production than consumer applications.

He said that most generative video companies are consumer-focused today. “They build simple models, where you can prompt a chatbot and generate some clips with cool effects. Our focus is on the technology required by Hollywood studios. What are the requirements for major brands to create Super Bowl commercials?” Layer-based editing – Marey, unlike other generative video models allows users to edit the foreground separately from the midground, allowing them more control over video composition.

Talukdar explained that the model is heavily based on controllability. “You need to be able to control the output more — to be able to change characters. It’s the only model that allows for layer-based editing. You can edit the background, mid-ground, and foreground separately. It’s also the first model built for Hollywood, purpose-built for production.”

In addition, he told VentureBeat that Marey relies on a diffusion-transformer hybrid model that combines diffusion and transformer-based architectures.

“The models are diffusion-transformer models, so it’s the transformer architecture, and then you have diffusion as part of the layers,” Talukdar said. “Controllability is usually introduced through the layers.”

Moonvalley has been funded by VCs of renown, but not at the same level as other AI video startups.

This week, Moonvalley announced a seed round of $70 million led by Bessemer Venture Partners. Khosla Ventures. and General Catalyst. Investors Hemant Taneja and Samir Kaul have also joined the board of directors. Talukdar (19659036) noted that Moonvalley has received substantially less funding than some of its competitors — Runway, for example, is reported to be having raised approximately $1.5 billion. The company has assembled an elite team of AI engineers and researchers to optimize its resources.

He said that the company raised $70 million. “We raised a lot less than our competitors,” he added. “But it really boils down the team — having the ability to build that architecture more efficiently, compute and all those other things.” Moonvalley will gradually expand access in the coming weeks. Talukdar confirmed that “Right Now, there are a number studios who have access to it and we have a beta group with a few dozen filmmakers.”

The hope is that the software will be available in a few weeks, or at worst, within a few months. But with AI video startups such as Runway and Pika, they are not alone. Honor is a field that is becoming more competitive as it continues to add new features such as character voice and movement.