WarpHammer: Training-Free Framework Tackles Extreme Novel View Synthesis with 3D Object Priors

WarpHammer is a training-free novel view synthesis framework that addresses a persistent failure mode in projection-conditioned NVS pipelines: when viewpoint changes are large (beyond roughly 110°), warped scene representations become sparse and riddled with mirror-like artifacts, causing generators to lose both pixel content and camera cues. WarpHammer resolves this by augmenting the warped scene with an explicit 3D reconstruction of the foreground object, sourced from a native 3D generative prior such as SAM3D, densifying exactly the region where degradation is worst — without fine-tuning the underlying video diffusion model.

A notable additional capability is cross-instance object fusion: WarpHammer can incorporate an auxiliary image of a different instance of the same object category (e.g., a casual snapshot of a car paired with a manufacturer studio shot) without requiring user-supplied camera poses. A pretrained multi-view geometry foundation model processes the reference and auxiliary images jointly to predict a unified point cloud, yielding richer geometry than single-image reconstruction alone. Across five benchmarks, the method is reported to maintain stable novel views at viewpoint deviations where strong baselines such as GEN3C collapse, and is described as the first scene-level NVS method to support pose-unknown auxiliary object views from external sources.