Tencent Open-Sources SCoPE: 3D-Consistent Video Generation from a Single Image, Errors Down 40%+

camenduru · x · 2026-08-14

Tencent's ARC team has open-sourced SCoPE (Sightline-Coordinate Positional Encoding), a video diffusion transformer model. It generates videos from a single still image and a predefined camera trajectory, requiring no game engine or 3D reconstruction.

Key Technical Highlights:

Sightline-Coordinate Positional Encoding: Injects camera ray information into tokens as positional coordinates, giving the model spatial awareness before generating any pixels.
Inherent 3D Consistency: With geometry embedded in the coordinates, looping paths naturally return to the starting scene, achieving 3D consistency for free.
Minimal Parameter Overhead: Adds new elements only into the Q and K of the attention mechanism, accounting for less than 0.1% of new parameters.
Performance Boost: On a 14B backbone, it reduces camera rotation error by 29% and translation error by 43%.

Original post →

Tencent Open-Sources SCoPE: 3D-Consistent Video Generation from a Single Image, Errors Down 40%+

More from Multimodal