RoGe merges reconstruction and generation without a lossy intermediate step.

Creating a smooth video from just a few photos is tricky. A computer needs accurate 3D geometry from the photos it has, plus smart guesses to fill in unseen spots. Older systems tried to solve this by creating explicit 3D models or intermediate rendered pictures. But these stepping stones lose details, pass along mistakes, and do not let the image generator help fix the 3D shape. Researchers built RoGe to fix this gap. RoGe creates an internal, implicit representation of the scene from sparse input views. It shoots virtual camera rays into this representation to pull out geometric features along a chosen camera path. A video diffusion modelvideo diffusion modelAn AI that generates videos by learning patterns from data. takes these features directly, skipping any 3D middleman. Because both parts train together, the image generator helps refine the underlying geometry. RoGe created higher-quality images and smoother videos across the camera path than previous baseline methods.

Why it matters

For creators building virtual worlds or videos, this method produces more realistic and stable scenes from just a few images. It means better quality content without needing extensive data or manual fixes.

Questions answered