A video model can preserve a subject’s visible details yet lose its shape as soon as the camera reveals a new side. The key idea here is to give the generator a separate, color-independent cue for geometry: a turntable of surface-normal maps reconstructed from the reference views. This targets cross-view consistency directly rather than asking appearance references alone to imply a coherent 3D object.
Key Findings
- The method reconstructs a subject from four views and renders 25 normal-map views around a yaw orbit. Because the geometry cue is provided as context rather than aligned frame by frame, it can guide camera changes without dictating the generated video’s timing.
- Its automated synthesis pipeline renders geometry-grounded training videos across 12 camera motions and 88 trajectory configurations. Showcase-4K supplies this training data, while ShowcaseBench evaluates on held-out assets with known meshes and camera parameters.
- On ShowcaseBench and the real scanned-object set NAVI, the paper reports higher identity similarity and lower 3D shape error than the compared methods. Removing the geometry condition raises GT-CD from 0.0824 to 0.2092 on ShowcaseBench and from 0.0975 to 0.1987 on NAVI, indicating that the geometry cue contributes materially to shape consistency.
- The gains are not uniform across every quality measure: aesthetic scores trail the best baseline on both sets. A small human comparison study favors the method, but its 24 evaluation videos make it supporting evidence rather than a broad user study.
Method and Tradeoffs
The reference images are used to estimate depth and 3D points, which are fused into a 3D Gaussian representation. Rendering that reconstruction from multiple angles yields normal maps that describe surface direction without introducing competing color or texture information. The normal sequence is supplied alongside the reference images to a pretrained image-to-video model.
This approach is best suited to mostly rigid subjects whose reference views depict a compatible shape. It is less reliable when images show different poses that cannot be reconciled into one static reconstruction, or when realistic interaction between the subject and its background matters: the geometry cue describes the subject, not a physically rendered scene.
