SCoPE Demo: 3D Video from Single Image Without Game Engine

yshan2u · x · 2026-08-17

SCoPE released a live demo generating videos from a single still image along predetermined camera paths (S-curves, orbits, backtracks) without a game engine or 3D reconstruction. The key innovation is replacing standard positional encoding (tensor coordinates like "row 12, col 30") with camera rays for each token. Since rays belong to the camera, the model establishes geometric coordinates before pixel generation, achieving 3D spatial understanding within a 2D workflow.

Original post →

More from Multimodal

Multimodal channel →