UniWorld-View: Large-Baseline Novel View Synthesis via Video Diffusion

Haiyang Zhou · hf · 2026-08-06

UniWorld-View introduces a unified framework for controllable large-baseline novel view synthesis from monocular inputs.

The Problem: Reconstruction-based approaches (like NeRF and 3DGS) deteriorate under sparse inputs, while existing generative methods struggle with large-baseline view synthesis due to inaccurate geometric guidance.
The Method: The framework integrates explicit 3D guidance with generative diffusion modeling. It uses an occlusion-aware point cloud rendering strategy to resolve visibility ambiguities and provide accurate priors for diffusion-based synthesis.
Performance: By coupling this rendering strategy with powerful video diffusion backbones, it achieves high-fidelity novel view generation even under extreme camera motions and wide-baseline changes. It can also provide multi-view videos for downstream dynamic 3DGS reconstruction.
Evaluation: Experiments on the WorldScore benchmark and zero-shot NVS benchmarks demonstrate its effectiveness in controllability, geometric consistency, and visual fidelity.

More from Multimodal