Video Rebirth Raises $80M to Build Real-Time World Model AI

Startup challenges tech giants with efficient AI video model

Video Rebirth, a Singapore-based AI video startup barely two years old, has secured $80 million in seed funding from investors including AMD Ventures and Hyundai Motor's ZER01NE to pursue an ambitious goal: building a world model that can generate interactive, physics-accurate 3D environments in real time.

The company's flagship product, Bach, ranks sixth on an Artificial Analysis text-to-video leaderboard, trailing only models from Alibaba, ByteDance, Kuaishou Technology, and xAI. Bach holds the distinction of being the highest-ranking startup model and offers the lowest cost per minute of video generation among the top ten competitors.

Founded by Liu Wei, formerly a distinguished scientist at Tencent who led development of the company's Hunyuan AI model, Video Rebirth operates with just 30 employees across Singapore and Hong Kong. The startup closed its March seed round with backing from AMD Ventures, Hyundai Motor Group's venture arm, Korean conglomerate CJ Group's Hiven, game developer Actoz Soft, Qiming Venture Partners, and Gaw Capital. Video Rebirth plans to raise additional funding in July, according to details first reported by Forbes.

Technical efficiency drives competitive positioning

Liu attributes Video Rebirth's ability to compete against well-capitalized rivals to proprietary technology that reduces both training and inference costs. The company's multi-step sampling loss technique trains models to anticipate and correct errors during generation, requiring fewer computational steps than traditional approaches. Bach generates videos using what Liu describes as "a fraction of" the budget required by comparable frontier models, though he declined to specify exact figures.

The startup trains on fewer, higher-quality videos—including licensed movies, music videos, and in-house footage at 720p resolution. Bach's architecture splits prompt adherence and visual generation into separate tasks, unlike competitors that use a single system for both functions. This division of labor improves computational efficiency, Liu explained.

Bach can generate multi-shot videos up to 45 seconds long from reference images and text prompts, compared to ByteDance's Seedance 2.0, which caps at 15 seconds. The model targets enterprise clients in advertising, entertainment, filmmaking, and gaming, with particular strength in maintaining product consistency for e-commerce and generating facial expressions for film production.

World model ambitions target autonomous systems

Video generation serves as a stepping stone to Video Rebirth's primary objective: a world model called Olympus, planned for release by the end of 2026. Unlike traditional 3D simulations that require pre-programmed responses, world models use AI to understand physical properties and simulate outcomes in novel situations.

Olympus will function similarly to Google's Genie 3, allowing users to navigate generated environments, but will add environmental sound generation including collision effects and footsteps. Hyundai Motor, a major investor, sees potential applications in training physical AI for autonomous driving and robotics. The automaker owns Boston Dynamics, a leading robotics company.

Why it matters

World models represent a potential breakthrough for autonomous vehicles, robotics, and game development, but remain in early stages. Video Rebirth's approach—building from video generation toward full physical simulation—offers a path to commercially viable applications if the technology can achieve real-time performance at scale. The startup's ability to compete on technical benchmarks while maintaining cost efficiency demonstrates that smaller, focused teams can challenge tech giants in capital-intensive AI domains. However, the timeline to mass adoption remains uncertain, with Liu acknowledging that world models will likely take longer to reach mainstream use than large language models did.

"We do video generation in order to build a world model," Liu stated. "In three years, we'll prove that the physical world can be simulated in real time."

Competing world model efforts include projects from Google, Nvidia, OpenAI, Google-backed Runway, and World Labs, cofounded by AI pioneer Fei-Fei Li. Alibaba and other tech giants are also pursuing the technology.

These details were first reported by Zinnia Lee for Forbes.