The path to truly interactive AI requires addressing three major hurdles: latency, consistency, and quality. By acquiring the team and technology of Morpheus AI and leveraging the expertise of the founders of Dynamics Lab and Lucid AI, Roblox is one step closer to realizing its interactive AI vision. For too long, video models have been locked into slow, bidirectional, offline generation. Researcher Xun Huang made the breakthrough and invention through years of post-training experimentation. self-enforcingan innovative recipe for converting slow offline video models into fast, autoregressive interactive generation engines. Armed with this breakthrough, Huang founded Morpheus AI in the relentless pursuit of real-time responsiveness. This key technology serves as the foundation built by a very strong team of researchers at Morpheus AI.
The original goal of Morpheus AI was to train egocentric video world models in a purely pixel latent space to simulate real-world mechanics and physics, and its ultimate realization is in Roblox Reality. At Roblox, we realized that the standalone world model, while powerful, fundamentally lacked the long-term memory, consistent logic, and structured user input needed for a massively multiplayer platform. This is where integration with Roblox’s engine is a clear opportunity. Our goal is for Morpheus AI’s models to accelerate the evolution of Roblox Reality’s Roblox Video Model (Super Upsampler) technology, overlaying visual components such as higher-fidelity textures, realistic lighting, and fluid physics, while the Roblox Engine manages structured symbolic logic and multiplayer state synchronization.
But Morpheus and Huang’s groundbreaking work is just one element of broader AI innovation at Roblox.
In late 2025, Dynamics Lab founder Joe Chen partnered with Roblox. At Dynamics Lab, Chen and his team focused on building a powerful, flexible, and responsive real-time general-purpose domain generation world engine. Users could upload any image, including photos, concept art, classic paintings, and children’s drawings, and step into the world as a live, interactive world. Users can direct text to create surreal scenes and wild events in their world, then share the link so friends can join and play right away.
In November 2025, Lucid AI founder Alberto Hojel also joined Roblox. Hojel founded Lucid AI on the following principles: A world generated by AI It needed to be reliable and consistent, with a balance between stunning visuals and definitive game logic. Hojel realized that while video world models can produce impressive, physics-rich realities, they fundamentally lack the long-term states and predictable progression required for true games. Based on this founding principle, Hojel invented the Roblox game cartridge harness. This augments the video world model with the reliable, deterministic logic of a dedicated game engine like Roblox. This critical convergence transforms a mere explorable world into a high-fidelity, interactive, multiplayer gaming platform.
The goal of the Roblox Reality architecture to achieve 2K resolution at 60 Hz is an unprecedented innovation for video world models. This is why deep expertise in optimization and KV compression is so important. Move the Roblox video model (Super Upsampler) to an adjacent edge data center with H200/B200 class GPUs. For example, research by Morpheus AI pioneered long context capabilities in world models. This is essential to maintaining visual consistency across the immersive experiences Roblox creators build.
Integration between game engines and video models is the missing piece to achieving a consistent, photorealistic, multi-user world. This hybrid architecture combines the structured simulation power of the Roblox engine with the generative vision of the Morpheus AI world model. Building a team of innovative founders creates a talent-rich environment that accelerates innovation and streamlines the development of high-fidelity interactive world creation.
We are very excited to have these leading AI innovators partner with Roblox. We are poised to leverage technologies such as self-forcing and pixel latent world models to build the foundational layer for photorealistic multiplayer games. Please pay attention to the moment of change that will occur in the future.
