Join our daily and weekly newsletters for the latest updates and the exclusive content on AI coverage. Learn more
A few years ago, there was nothing like a “video generator model on AI”.
Today, there are dozens, including many capable of making an ultra-high definition video, ultra-realistic Hollywood caliber in seconds from text or images promoted by the user and existing video clips. If you have read VentureBeat in recent months, you have undoubtedly encountered articles on these models and companies behind them, Runway Gen-3 has Google Veo 2 To be opened for a long time but ultimately available Sora has Luma Ai,, Pikaand Chinese smugglers Kling And Hail. Even Alibaba And a startup called Genmo offered open source video models.
Already, these models have been used to make major blockbusters games, from Everything, everywhere at the same time has The real HBO detective: Night Country clips and television advertisements Toys r ‘us And Coca-Cola. But despite the relatively rapid hug of Hollywood and filmmakers, there is still a big problem for potential buildings: the concerns of copyright.
From the best we can say, given that most startups in the IA video model do not publicly share specific details of their training data, most are trained on large expanses of videos downloaded from the web or collected from other archive sources, including those with copyrights whose owners may or may not have really granted an express authorization to IA video companies to train. Actually, The track is one of the companies faced with a collective appeal (always making a way in court) on this very problem, and Nvidia would have scratched a huge band of YouTube videos So for this purpose. The dispute is in progress on the question of whether scratching data, including videos, constitutes fair and transformational use.
But now there is a new alternative for those who are concerned about copyright and did not want to use models when there is a question mark. A startup called Moonvalley – Founded by the ancient Google Deepamiders and researchers from Meta, Microsoft and Tiktok, among others – introduced Marey, a video model of the generative AI designed for Hollywood studios, filmmakers and business brands. Positioned as a basic video model of “clean” at the cutting edge of technology, Marey is formed exclusively on data possessed and approved, offering an ethical alternative to AI models developed using the scraped content.
“People said it was not technically possible to build a video AI video model without using scratched data,” said CEO and co -founder of Moonvalley, Naeem Talukdar, in a recent video call interview with Venturebeat. “We have proven the opposite.”
Marey, available now on the basis of a waiting list by invitation only, joined Adobe Firefly video modelWhat the software supplier established for a long time said is also of business quality – having been trained only on licensed data and Adobe Stock Data (to the dismay of certain contributors) – and provides compensation for companies for use. Moonvalley also provides compensation on Clause 7 of this documentSaying that he will defend his customers at his expense.
Moonvalley hopes that these features will make Marey attractive to the big studios – even if others such as The track agreements with them – and filmmakers, among the countless and constantly increasing new IA video creation options.
No more “ethical” video?
Marey is the result of a collaboration between Moonvalley and Asteria, a film and animated studio of AI led by artists. The model is designed to help rather than replace creation professionals, providing filmmakers with new tools for AI video production while maintaining traditional industry standards.
“Our conviction was that you will not get consumer adoption in this industry, unless you did with industry,” said Talukdar. “The industry was noisy and clear that so that they can really use these models, we must understand how to build a clean model. And until today, the upper track was that you couldn’t do it. »»
Rather than scratching the internet for content, Moonvalley has established direct relations with creators to concede to their images. The company has taken several months to establish these partnerships, ensuring that all the data used for training have been legally acquired and entirely under license.
Moonvalley’s license strategy is also designed to support content creators by compensating them for their contributions.
“Most of our relationships have really come down now that people have started to hear what we are doing,” said Talukdar. “For creators of small towns, a large part of their images is just sitting. We want to help them monetize it and we want to make artists focused. It ends up being a very good relationship. »»
Talukdar told VentureBeat only if the company assesses and revises its remuneration models, it generally compensates for creators according to the duration of their images, paying them an hourly or carefully rate under fixed -term license agreements (for example, 12 or four months). This allows potential recurring payments if the content continues to be used.
The company’s objective is to make high -end video production more accessible and more profitable, allowing filmmakers, studios and advertisers to explore the narration generated by AI without legal or ethical concerns.
More cinematographic control – beyond the text guests, images and camera directions
Talukdar explained that Moonvalley has adopted a different approach with its Marey AI video model than existing AI video models by focusing on professional quality production rather than consumption applications.
“Most generative video companies are now more focused on consumers,” he said. “They build simple models where you cause a chatbot, generate clips and add nice effects. Our goal is different: what is the technology necessary for Hollywood studios? What do major brands need to make Super Bowl advertisements? »»
Marey presents several progress in the video generated by AI-AI, in particular:
- Native HD generation – generates a high definition video without counting on scaling, reducing visual artefacts
- Extended video length – Unlike most AI video models, which only generate a few seconds of sequences, Marey can create 30 seconds sequences in a single pass.
- Edition based on layers – Unlike other generative video models, Marey allows users to modify the first plan separately, the middle and the background, offering more precise control over the video composition.
- Storyboard and entrances based on sketches – Instead of relying solely on text prompts (as many AI models do), Marey allows filmmakers to create with storyboards, sketches and even live references, which makes it more intuitive for professionals.
- More sensitive to packaging entries – The model was designed to better interpret external entries such as motion drawings and references, which makes the video generated by AI-Plus controllable.
- “Netifice” video editor – Moonvalley is developing companion software for Marey, which works as a generator-native video editing tool that helps users more effectively manage projects and deadlines.
“The model itself is just built very strongly around controllability,” said Talukdar. “You must have many more controls around the output – be able to change the characters. It is the first model that allows you to make edicts based on layers, so you can modify the first plan, the middle and the background separately. It is also the first model built for Hollywood, built for production purposes. »»
In addition, he told VentureBeat that Marey relies on a hybrid model of diffusion transformer which combines architectures based on diffusion and transformers.
“The models are models of a diffusion transformer, so it’s the architecture of the transformer, then you have a distribution in the diapers,” said Talukdar. “When you introduce controllability, it is generally through these layers that you do.”
Funded by VCs of big names but not as much as the other video startups of the AI (still)
Moonvalley also announces this week a seed lap of $ 70 million led by Bessemer Venture Partners, Khosla Ventures and General Catalyst. Investors Hemant Taneja, Samir Kaul and Byron Deeter also joined the company’s board of directors.
Talukdar noted that MoonValley funding is significantly lower than some of its competitors, so far – the track would have lifted $ 270 million in total on several laps – But that the company has optimized its resources by assembling an elite team of IA researchers and engineers.
“We have collected about $ 70 million, a little less than our competitors, certainly,” he said. “But it really comes down to the team – having a team that can build this architecture much more effectively, calculate and all these different things.”
Marey is currently in a limited access phase, with certain studios and filmmakers testing the model. MoonValley plans to gradually expand access in the coming weeks.
“Right now, there are a number of studios that have access to it, and we have an Alpha group with a few dozen filmmakers,” confirmed Talukdar. “Hope is that it will be entirely available in a few weeks, worse cases in a few months.”
With the launch of Marey, Moonvalley and Asteria aim to position themselves at the forefront of the production of films assisted by AI, offering studios and brands a solution that integrates AI without compromising creative integrity. But with video startups on AI Hède Continuing to add new features such as voice and character movements, the field becomes more and more competitive.
