Camera-Conditioned Video Generation

The field studies how to generate video that follows a requested camera motion while keeping the scene coherent across changing viewpoints. Common approaches condition diffusion models or video transformers on camera poses, depth, point tracks, or rendered 3D structure; some also use persistent scene representations or memory for long sequences. These methods support novel-view synthesis and interactive world models, with evaluation on camera-following, multi-view consistency, and long-horizon exploration tasks. A central challenge is balancing realistic visual detail with accurate geometry and stable scene identity, especially when views are sparse, objects move or become occluded, or generation must run in real time.