CubeComposer turns a single clip into a 4K 360-degree VR video

CubeComposer is an AI model that converts a regular single-camera video into an explorable 360-degree clip. Simultaneously, it also upscales the clip to 4K in the process.

Interactive demos show scenes like snowy fields and highways where viewers can rotate the camera freely. However, CubeComposer hallucinates plausible content even behind the original viewpoint.

Compared with previous systems like Argus and Viewpoint, CubeComposer delivers far better quality. Competing outputs are heavily distorted and blurry, while CubeComposer’s reconstructions are relatively clean.

The model uses a diffusion architecture that decomposes the input into six cube faces. In addition, it generates each segment with sparse-attention context pooling, blending them with cube-aware positional encoding and padding for seamless transitions.

The team has open-sourced the code via GitHub, including local install instructions for running on your own GPU. For creators building VR content from standard footage, this offers a new path without needing multi-camera rigs.