Generative View Stitching

The paper proposes Generative View Stitching (GVS), a sampling algorithm that enables parallel generation of entire video sequences conditioned on predefined camera trajectories, addressing collision and autoregression collapse issues in camera-guided video synthesis. It resides in the Explicit Camera Pose Conditioning leaf, which contains six papers including CameraCtrl, MotionCtrl, and CamViG. This leaf sits within the broader Camera Trajectory Parameterization and Control Mechanisms branch, indicating a moderately populated research direction focused on direct pose-based conditioning rather than implicit motion encodings.

The taxonomy reveals neighboring leaves addressing related but distinct challenges: Implicit Motion Representations (four papers using optical flow or learned encodings) and Training-Free and Plug-and-Play Control (two papers enabling inference-time guidance without fine-tuning). The paper's emphasis on parallel sampling and stitching connects it to Multi-View and 3D-Consistent Generation themes, particularly the Multi-Camera Synchronized Generation and Depth-Guided methods that also prioritize geometric coherence. The scope notes clarify that explicit pose conditioning excludes trajectory-based or flow-driven approaches, positioning GVS firmly within methods that directly parameterize camera extrinsics.

Among the three contributions analyzed, the GVS sampling algorithm was examined against zero candidates, while OmniGuidance (bidirectional temporal conditioning) was compared to ten candidates with none providing clear refutation, and the loop-closing mechanism was assessed against four candidates, also without refutation. The total search examined fourteen candidates across all contributions, a limited scope that captures immediate neighbors but does not constitute exhaustive coverage. The absence of refutable prior work among these candidates suggests that the specific combination of parallel stitching, bidirectional conditioning, and loop-closing for camera-guided generation may be relatively unexplored within the examined literature.

Based on the limited search scope of fourteen candidates, the work appears to introduce a distinct approach within its taxonomy leaf, particularly in addressing autoregression collapse through parallel sampling and future conditioning. However, the analysis does not cover the full fifty-paper taxonomy or broader diffusion stitching literature beyond top semantic matches. The novelty assessment is therefore constrained to the examined subset, and a more comprehensive search might reveal additional overlapping methods in related diffusion planning or multi-view synthesis domains.