A “clean” image and video generation model, trained exclusively on “ethically sourced” data, is being positioned for enterprise-grade productions in Hollywood next year as an alternative to offerings from tech giants accused of data scraping.
The new model, dubbed “Marey,” after early cinema pioneer Étienne-Jules Marey, will be available in January or February 2025. As an exclusive industry survey indicated earlier this month—the latest in a series published by Variety Intelligence Platform in recent years—the entertainment business is currently in a heavy exploratory phase regarding AI technologies.
The model comes from Los Angeles-based generative video AI research startup Moonvalley in partnership with the “artist-led” AI film and animation studio Asteria Film, owned by the nonfiction film and television studio XTR. Developed in close collaboration over the past six months, the two companies will jointly own and control the model.
Popular on Variety
For Moonvalley and Asteria, the most important factor differentiating the Marey model versus other video models in the market is that it has been trained exclusively with data provided with the permission and payment of content owners.
VIP+ Special Report: Generative AI & Licensing
The vast majority of the dataset being used to train the Marey model consists of content Moonvalley licensed under multiyear deals from various data provider partners, primarily from creators who opted in and selectively provided their data. For the remaining portion of the dataset, Moonvalley purchased the rights to small, specialized datasets and even enlisted the help of the production team at Asteria to create brand-new footage in its studio in order to fill in specific gaps missing from the dataset.
By contrast, most generative AI companies train models on “publicly available” data scraped from the web, some of which is copyrighted, which they defend as fair use under U.S. copyright law. This is true of some of the leading video generation models.
OpenAI, which released its awaited AI video tool Sora last week, has said it would be “impossible” to train AI models without using copyrighted material. Ex-OpenAI CTO Mira Murati didn’t deny in an interview that Sora was trained on YouTube videos. Meta trained its latest video model Movie Gen on a “combination of licensed and publicly available datasets,” according to a blog post.
A 404 Media report suggested generative video startup Runway trained its video model on thousands of videos from the YouTube channels of entertainment companies Netflix, Disney, Nintendo and major social media creators.
In order to make generative AI “usable” in Hollywood, Mooser and Taldukar argue, major studios, artists filmmakers need an AI model to be clear of the legal complications and ethical harms that have emerged directly from just these kinds of unlicensed use of copyrighted data to train AI models.
“AI is a powerful tool for the future of Hollywood, but it’s not usable until a clean model is built that’s not trained on other people’s stolen data,” said Bryn Mooser, CEO of XTR and Asteria. Mooser is an Oscar-nominated filmmaker and previously co-founded the immersive entertainment studio RYOT, acquired in 2016 by HuffPost.
“In order for creators to be able to use and get behind generative AI, it has to be trained and built in a way that respects the creators and the industry that it’s coming into,” Moonvalley co-founder and CEO Naeem Talukdar told Variety. “The problem with ignoring or being lackadaisical about where you’re actually sourcing the data from is that the people that you’re alienating the most are the people that you need to be building this technology for.”
Initially, the companies’ Marey model will principally be used in workflows by artists and filmmakers at Asteria to produce projects across animation, fiction and nonfiction film and television.
Asteria and Moonvalley also plan to offer Marey as an enterprise software solution, and Asteria is engaged in early discussions with major and independent studios in Hollywood about partnerships in which Asteria and Moonvalley would be able to develop bespoke versions of the tool and advise on production workflows to meet the needs of specific studio projects.
Similar to features available in other generative video software, Marey will provide capabilities, including text-to-video, image- and video-to-video and custom model creation (aka fine-tuning). Moonvalley also intends to build granular creative control capabilities into the model, which Talukdar said would be the team’s biggest focus through 2025 once the model has been fully trained and released.
“In order for generative models to become useful for creative teams or for studios, they should be able to control a model with close to the same level of accuracy and precision as they would have over a production set, the camera, lighting, or scene composition,” said Talukdar.
VIP+ Analysis: Meet the New Breed of Studios Pioneering Gen AI-Powered Production
Moonvalley and Asteria still face competition as video generation models flood into the market from OpenAI and startups such as Runway, Luma Labs, Black Forest Labs, Pika and Stability AI, not to mention numerous contenders out of China. Tech giants Meta and Google are developing their own video models, Movie Gen and Veo 2, respectively, though neither has been publicly released.
Meanwhile, Adobe is taking a similar approach as the Marey model with its Firefly video model. Currently available in limited public beta, Adobe has described its video model as “commercially safe” for enterprise and professional customers due to only training on content the company has permission to use.
Several of these AI companies are also aspiring to work with Hollywood studios and filmmakers, some of which have taken early steps to work with them. In September, Lionsgate entered a partnership with Runway to develop a jointly managed video model exclusively for internal use by Lionsgate and its filmmakers, trained on some of the film studio’s own catalog.
Other major studios are having similar conversations behind the scenes. “Avatar” director James Cameron joined the board of directors of at Stability AI. Horror studio Blumhouse participated in a pilot program for Meta’s Movie Gen.
The partnership announcement comes on the heels of Asteria Film Co. being formally established under nonfiction studio XTR, following its acquisition of the AI animation studio Late Night Labs in September. It also follows Moonvalley announcing a $70 million funding round in November, led by General Catalyst and Khosla Ventures, with participation from Bessemer Ventures.
Moonvalley has been focused on building video generation models for creatives with a current team of about 30 generative AI researchers who previously worked at DeepMind, Meta, TikTok, and Microsoft. The Moonvalley founding team also includes Mooser, who heads a creative team of 30 at Asteria (pictured above, including Moonvalley researchers), working out of the Mack Sennett soundstage in Los Angeles.
Variety VIP+ Explores the Development of Gen AI From All Angles — Pick a Story