MetaView uses diffusion and depth to synthesize views from a single image
Diffusion model handles large viewpoint changes by blending implicit geometry with metric depth cues.
Current visual generation models produce high-quality content but lack coherent spatial structure. Existing novel view synthesis methods either enforce strict geometry consistency (limiting large view changes) or use implicit scene modeling (sacrificing camera control). MetaView bridges this gap with a diffusion-based approach that introduces implicit geometry priors from a feed-forward network, regularizing structure without a restrictive reconstruction pipeline. By also leveraging metric depth, it anchors generation to a metric scale, enabling both geometry consistency and precise controllability.
Extensive experiments show MetaView significantly outperforms existing methods on challenging monocular large viewpoint changes, demonstrating superior generalization. The framework allows rendering novel views from a single image with realistic spatial coherence. Accepted to ECCV 2026, the code is publicly available. This work represents a practical step toward flexible, controllable view synthesis for VR/AR and 3D content creation.
- Combines implicit geometry priors (from a feed-forward network) with explicit metric depth for structure regularization.
- Enables large viewpoint changes from a single image while maintaining geometry consistency and precise camera control.
- Accepted to ECCV 2026 and outperforms existing methods on monocular view synthesis benchmarks.
Why It Matters
MetaView unlocks practical single-image 3D scene exploration, accelerating VR/AR and creative workflows.