After halting development of an NFT-based video game, Seattle-area startup Irreverent Labs is attracting investor attention for its new focus: an artificial intelligence model that can generate video.
The company announced Thursday it raised an undisclosed amount of funding from Samsung NEXT, the Silicon Valley-based venture capital arm of Samsung Electronics.
Irreverent aims to develop a video foundational model and API that creators, game studios, and developers can use to generate content.
It is among a wave of startups and large tech firms developing tools that generate videos from prompts in natural language, images, or brief video snippets.
Founded in 2021, Irreverent first emerged with the release of MechaFightClub, a play-to-earn video game that blended blockchain technology and AI. Players could trade, sell, and fight robotic chickens in order to earn what it called “nuggets,” a form of digital currency that could be spent on in-game purchases or withdrawn for use on the wider Solana blockchain.
Irreverent raised $45 million from Andreessen Horowitz and others to launch the game. However, just a year later, the startup decided to shut it down, citing shifting U.S. regulatory posture toward cryptocurrency.
“The decision to pause the development of the game and focus on the foundation model was a big decision for us as a team,” co-founder and CEO Rahul Sood told GeekWire. “But we believe it was the right decision.”
Sood said Irreverent used its own AI tools to develop the game. In the process of animating a fight scene, a researcher discovered a way to generalize the solution for broad video applications.
“It should be very difficult for people to look at what we have as an output and say, ‘Oh, that’s AI-generated,'” CTO David Raskino said.
“It was a breakthrough,” he said.
After the discovery, Irreverent fully embraced AI video generation, tapping into its remaining cash runway to purchase GPUs to store data for the video foundation model, Sood said.
One of the biggest challenges in creating AI-generated video is what is referred to as “temporal consistency,” or maintaining smooth and natural transitions between frames, said David Raskino, CTO and co-founder of Irreverent Labs. With conventional models, pixels get “confused about which object they belong to,” resulting in poor quality video outputs, he said.
The startup aims to solve for these challenges by using a hybrid approach, combining video analysis, synthesis and entropy models to create cohesive and realistic videos.
“It should be very difficult for people to look at what we have as an output and say, ‘Oh, that’s AI-generated,'” Raskino said.
Irreverent’s partnership with Samsung could eventually lead to a distribution deal where its video foundation model gets integrated into the company’s devices, Sood said. The goal would be for a user to take a photo, then use the model to create a short-form video, right on their phone, he said.
The company also plans to develop an API, where developers can create applications on top of the model, he said.
“The possibilities that Irreverent Labs’ technology unlocks are vast, and the potential impact on the mobile devices in our pockets and backpacks to the televisions mounted on our living room walls is immense,” Samsung Next investor Joan Kim wrote in a blog post announcing the deal.
The emergence of AI has been a cause for concern over intellectual property rights, as content creators and studios fear their work is being used to train AI models without compensation or acknowledgment. There has been misuse over the creation of deep-fakes, where audio, images and videos are digitally altered to resemble the likeness of another person, often without their consent. AI is at the center of the Hollywood strike; actors and writers have expressed their concerns over AI being used to replicate their likeness, taking away job opportunities.
Irreverent is creating guardrails to “fine-tune” outputs based on specific customer needs, Sood said. For instance, a movie studio could adjust the settings to align with its own character IP, allowing them to monetize their video outputs, he said.
Irreverent is training its AI model on “everything that’s publicly available on the internet,” Sood said.
The Seattle startup is among a bevy of companies trying to roll out AI video generators. Runway ML, which recently raised $141 million, lets users create content through natural language prompts. Mobius AI, founded by former Google employees and reportedly valued at over $100 million, is also creating video AI tools. Adobe Firefly is testing an AI generator focused on video.
“The difference will be evident in the quality of what we’re producing,” Sood said of competition. “We believe in order to produce something that’s commercially viable, it needs to be really high quality and it needs to have some character consistency.”
He said the company has also invested heavily in its talent. Its distinguished scientist is Patrice Simard, the former chief scientist at Microsoft Research. The startup has 28 employees, and about half the team are former scientists.
Sood previously co-founded esports betting startup Unikrn, which was acquired in 2021, and led Microsoft’s venture capital arm. Raskino helped lead the VC fund with Sood and most recently oversaw a new security team at Microsoft.
