Creating a new view of a scene from a photo is hard. To judge quality, you compare with a known view. But each method may reconstruct the scene in a shifted coordinate frame, unfairly lowering its score.
The fix uses standard image matching. Match features between generated and real views, estimate camera pose by linking image points to reconstruction points, then refine the pose. Score only regions both views actually see.
This changes the leaderboard. A method's score rose from 14.52 to 22.04 decibels, and explicit Gaussian models (point-cloud-like reconstructions) overtook pure image-based video generators. A baseline using off-the-shelf video and geometry tools matched state-of-the-art without task-specific training.