In a landscape dominated by heavyweight competitors, Video Rebirth appears to be an unlikely contender in the high-stakes realm of AI video technology. Founded less than two years ago, this fledgling startup has already secured $80 million in funding and employed a lean team of 30 professionals based in Singapore and Hong Kong. Given that training state-of-the-art video models can cost tens of millions, one might think Video Rebirth would struggle to find its place in the competitive market.
However, just ahead of its flagship model's launch in May, Video Rebirth made an impressive entry into a benchmark leaderboard for AI-generated video, ranking sixth among industry titans like Alibaba, ByteDance, Kuaishou Technology, and xAI. Its flagship product, the Bach model, achieved this ranking as the top startup entry, boasting the lowest cost per minute of generated video among the leading competitors.
Liu Wei, cofounder and CEO of Video Rebirth, expressed that this achievement was a significant indicator of their innovative architectural approach. Liu envisions that creating AI-powered video engines is merely the initial phase. He aims to develop models capable of producing hyper-realistic visuals that adhere to the laws of physics, anticipating a future where these technologies could transform industries such as autonomous driving, robotics, and gaming. Liu is committed to crafting a "truly meaningful" world model that simulates understanding and reasoning akin to human instincts.
"We engage in video generation to develop a world model," Liu explains. "In three years, we aim to demonstrate that real-time simulation of the physical world is feasible."
In March, Video Rebirth completed an $80 million seed funding round, with investments from notable firms such as AMD Ventures, Hyundai’s ZER01NE, Hiven, Actoz Soft, Qiming Venture Partners, and Gaw Capital. Though the company is planning another funding round in July, specifics remain under wraps.
Fang Wei, a senior investment manager at Hyundai Cradle, articulated the belief that video generation serves as a pathway toward building comprehensive world models, aligning with Video Rebirth’s vision to drive revolutionary advancements in physical AI.
The Bach model is targeted at enterprise clients in diverse sectors including advertising, entertainment, and gaming. It features the capability to generate multi-shot videos lasting up to 45 seconds from reference images and text prompts—substantially outperforming competitors like ByteDance’s Seedance 2.0, which limits outputs to 15 seconds.
Operating in a demanding and costly environment, Video Rebirth distinguishes itself through its innovative technology, reducing inference costs significantly. Liu claims that Bach’s generation of a 10-second clip is far more efficient compared to other prominent models, although specific figures were not disclosed. This efficiency is attributed to a proprietary technique called multi-step sampling loss, which enhances the model’s ability to foretell and rectify potential errors during video creation, thus expediting the process.
Moreover, Video Rebirth's approach to training uses fewer but higher-quality video datasets, including licensed film content and in-house productions, allowing for substantial savings compared to other advanced models. Bach’s architecture divides tasks related to visual generation and prompt adherence for improved computational efficiency.
In contrast, OpenAI faced substantial operational costs, as evidenced by its decision to discontinue its Sora platform despite substantial user engagement and previous agreements with major companies. Liu remarked that such financial pressures hindered OpenAI's potential, highlighting Video Rebirth’s more cost-effective strategy.
Beyond cost management, Video Rebirth excels in producing videos that accurately depict physical interactions, such as gravity and lighting, which have often been challenging in AI-generated content. Liu emphasized the importance of consistency for e-commerce applications and noted the model's proficiency in crafting detailed facial expressions and scenic visuals, attracting interest from various entertainment sectors.
According to Hyundai’s Fang, Video Rebirth’s emphasis on enterprise-grade reliability and precision addresses critical challenges in video generation, including the AI’s ability to comprehend cause-effect relationships in dynamic environments.
Looking toward the future, managing partner Alex Zhou of Qiming Venture Partners believes Video Rebirth has the potential to become a standard tool for professionals across multiple fields, akin to Adobe's role in the traditional creative industry. The startup is developing its world model, named Olympus, which aims to simulate interactive 3D environments and react intelligently to stimuli in real-time. This initiative aligns with a broader trend, as various firms explore world model capabilities for applications ranging from self-driving cars to advanced robotics.
World models are still in the early phases, capturing significant interest from major technology companies and well-funded startups. As industry experts weigh in, the consensus indicates that while world models represent a promising frontier in AI, they are yet to achieve the fidelity and cost-effectiveness required for widespread application.
Liu remains committed to advancing Video Rebirth’s vision, with plans for the Olympus model to launch by the end of 2026, designed to include reactive environmental sounds, enhancing the immersive experience. He believes that as their world model evolves and begins simulating complex physical scenarios in real time, it will unlock novel applications across various sectors beyond gaming.
Having previously held a prominent research position at Tencent, Liu recognized the potential in physical AI, inspiring him to launch Video Rebirth. With a strong team of experts, Liu is dedicated to pushing technological boundaries in pursuit of a commercially viable world model.
Despite the challenges that lie ahead, Liu expresses unwavering determination to advance this exciting field. “I will invest all my energy into research and development until we achieve a successful world model,” he asserts. “That breakthrough is on the horizon, without a doubt.”
