The widespread availability of tools to build generative AI has led to a Cambrian-like explosion of startups in this field. Abundant capital is not having an impact, nor is the falling cost of the necessary technology infrastructure.
In fact, one of the flashiest applications of generative AI, generative video, is in danger of becoming saturated. Labs like Genmo, Haiper, and Rhymes AI are rapidly releasing models that, in some cases, are nearly indistinguishable from previous state-of-the-art models.
Naeem Talukdar believes that what sets some generative video ventures apart from others is trust, rather than model ability. That's why he founded Moonvalley, a Los Angeles-based startup that is ostensibly developing more “transparent” generated video tools.
Talukdar led product growth at Zapier before founding Draft, a Y Combinator-backed company that hosts a marketplace for enterprise AI content. He hired Mateusz Malinowski and Mick Binkowski to launch Moonvalley. Both are former DeepMind scientists, where they worked on video generation technology.
“We shared the belief that video generation would transform media and entertainment, but the startups operating in this space did not have the characteristics needed to succeed,” said Talukdar. told TechCrunch. “Incumbents have been very hostile to artists, creators, and the broader industry.”
Talukdar points out that most generative AI companies train their models on public data, some of which is always copyrighted. These companies argue that the fair use doctrine protects this practice. For example, OpenAI claims that it can't properly train a model without copyrighted material, and Suno says that promiscuous training can be done by “listening to the genre and making yourself a rock song.” He claimed that he was no different from “a child who writes.”
Part of Moonvalley's founding team. From left to right: Mateusz Malinowski, Bryn Muser, Mikolaj Binkowski, John Thomas. Mikolaj Binkowski, John Thomas. Image credit:Moonvalley
But that hasn't stopped rights holders from filing complaints or seeking injunctions.
Vendors are getting very brazen even though lawsuits are piling up against them. Earlier this year, Mira Murati, former CTO of OpenAI, did not fully deny that OpenAI's video model Sora was trained on YouTube clips, in violation of YouTube's Acceptable Use Policy. It seems there are. Elsewhere, a report from 404 Media suggests that generative video startup Runway has harvested YouTube footage from channels belonging to creators like Disney and MKBHD without permission.
Canadian AI startup Viggle freely admits that it uses YouTube videos to power its video models. And like most of its rivals, it offers no recourse to creators whose work may have been washed away during training.
“Generative models must respect copyright, trademark, and publicity rights,” Talukdar said. “That's why we work so closely with the creators of our models.”
Moonvalley claims it doesn't have a fully trained video model yet, but is one of the few companies that uses exclusively licensed data from content owners who have “opted in.” To cover its bases, Moonvalley allows creators to request content be removed from their models, allows customers to remove their data at any time, and provides indemnification to protect users from claims of copyright infringement. We're going to provide you with a policy.
This is a bit of a pivot from Moonvalley's original pitch, which was a fairly general video generator for creating anime and fantasy-style clips. Moonvalley released two video generation models in beta to its strong Discord community of over 100,000 users before deciding to switch gears.
Output from one of Moonvalley's original video generator models. Image credit: Moon Valley
This new approach is similar to Adobe's approach to training Firefly video models using content licensed from the Adobe Stock contributor platform. Talukdar declined to say how much Moon Valley pays the creators of the clips, but it could be quite expensive. Bloomberg reported that Adobe is offering about $120 for every 40 to 45 minutes of video.
To be clear, Moonvalley does not source content itself. We work with anonymous partners who handle licensing agreements and package videos into datasets that Moonvalley purchases.
These partners, so-called “data brokers,” are in high demand these days thanks to the generative AI boom. The market for AI training data is expected to grow from around $2.5 billion today to nearly $30 billion within 10 years.
“We license high-quality data from multiple sources that work directly with creators and fully compensate them for the use of their content,” Talukdar added. “We ensure that we are using high-quality and diverse datasets.”
Unlike some “unfiltered” video models that simply insert human likenesses into their clips, Moonvalley is also committed to building guardrails around its creative tools. Similar to OpenAI's Sora, Moonvalley's model blocks certain content, such as NSFW phrases, and does not allow prompts to generate videos of specific people or celebrities.
Of course, no filter is perfect, but Talukdar says this “red teaming” is central to Moonvalley's release strategy.
“As the relationship between media and AI continues to rapidly evolve, and not without some skepticism, Moonvalley aims to establish itself as the most trusted partner for media organizations,” he said. Ta.
But can Moon Valley really compete?
As I alluded to earlier, companies like Google, Meta, and countless others are pursuing generative video with varying degrees of ethical considerations. Big tech companies are changing their terms of service to gain data advantages. Google trains its Veo video model on YouTube videos, while Meta trains its model on Instagram and Facebook content.
Moonvalley hopes to appeal to brands and creative houses, and some vendors are already making meaningful progress in this space. Runway recently signed a deal with Lionsgate to train custom models on the studio's movie catalog. Stability AI welcomes Avatar director James Cameron to its board of directors. OpenAI worked with brands and independent directors to demonstrate the potential of Sora.
And Adobe is targeting Moonvalley's target market: artists and content creators who want a “safer” (at least from a legal perspective) video production tool.
Moonvalley has three challenges. You need to convince your customers that your tools can compete with existing ones. We need to build enough runway to train and serve subsequent models. And they need to ensure they have a loyal customer base that won't switch to another provider anytime soon.
It's no surprise that many artists and creators are wary of generative AI, as it threatens to transform the film and television industry. A 2024 study commissioned by the Animation Guild, a union representing Hollywood animators and cartoonists, found that more than 100,000 U.S.-based film, television and animation jobs will be destroyed by AI by 2026. It is estimated that
“We are focused on building tools that empower creators to create more epic, immersive content than ever before,” when asked about the risk of creators losing their jobs due to generative AI. , Talukdar said.
On the runway front, Moonvalley's has made some progress. The company recently raised $70 million in a seed funding round co-led by General Catalyst and Khosla Ventures with participation from Bessemer Ventures. That will fund Moon Valley's research and development and employment.
The company currently has about 30 employees who previously worked at DeepMind, Meta, Microsoft and TikTok, Talukdar said.
“What sets us apart from other companies is our focus on the product,” he added. “While our core lies in training state-of-the-art generative models, we are also excited to build creative tools with advanced capabilities to turn these models into powerful instruments for professional creators, studios, and brands. We focus on that.”
Talukdar said he plans to release Moon Valley's first model later this year. If the company wants to beat out upcoming releases from Black Forest Labs, Luma Labs, Midjourney, and the elephant in the room, it'll need to hurry.