HuggingFace's Zuan Carreño has developed an early implementation of an interactive 360° AI-generated world using MiniMax H3. The project, which was demonstrated on YouTube, allows users to explore a generated environment with control over the viewport on mobile and desktop devices. The video, generated on a consumer Nvidia RTX 3080 Ti 16GB VRAM, has visible seams and resolution limitations, but it highlights the potential of immersive AI-generated content. The experiment uses a specialized equirectangular generation workflow with AI 360° prompting to create an interactive viewing experience.
The project is part of a broader effort to redefine how generative AI media is consumed. Unlike conventional video, which is predetermined, the equirectangular format allows users to choose where to look, changing the relationship between the viewer and the content. This format is particularly suited for immersive telepresence, robotics, and scene-centric world modeling. The equirectangular representation also has a 2:1 aspect ratio, which is essential for delivering panoramic video on platforms like YouTube and Facebook. The workflow uses a 2112 × 1056 resolution to maintain the equirectangular 360° geometry.
The technical challenge lies in ensuring the seamless connection between the left and right edges of an equirectangular image. Conventional neural video models do not automatically understand this requirement, leading to visible seams. The workflow addresses this by using specialized MiniMax H3 processing for equirectangular output, including an EquirectMiniMaxH3VAEDecode stage. The generated sequence is presented as an immersive environment where the viewer controls the viewing direction, allowing for interactive exploration.
Source: huggingface
