Scholarlink

We present NoPo3DFusion, a scalable novel view synthesis (NVS) framework that generates long, multi-view consistent video sequences from just two unposed RGB images with controllable camera trajectories. NoPo3DFusion combines NoPoSplat, a pose-free feed-forward 3D Gaussian Splatting (3DGS) based method designed for paired inputs, with CogVideo-X, a transformer-based video diffusion model. To enhance spatial consistency and guide the generative process, we introduce an observable mask as an additional condition for the diffusion model. We further propose an iterative inference strategy that enables efficient generation of long sequences while maintaining high visual fidelity and multi-view consistency. Experiments on RealEstate-10K and ACID show that NoPo3DFusion sets a new state of the art for feed-forward NVS from sparse and unposed images, highlighting the effectiveness of combining pose-free 3D reconstruction with video diffusion model. Code and models are available at https://github.com/maxShen701/NoPo3DFusion .

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

NoPo3DFusion: Scaling Two Images to Long Videos via 3D-Aware Iterative Diffusion

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xiaolong Shen,
  • Cheng Zhang,
  • Donny Y. Chen,
  • Tien-Tsin Wong,
  • Jianfei Cai

摘要

We present NoPo3DFusion, a scalable novel view synthesis (NVS) framework that generates long, multi-view consistent video sequences from just two unposed RGB images with controllable camera trajectories. NoPo3DFusion combines NoPoSplat, a pose-free feed-forward 3D Gaussian Splatting (3DGS) based method designed for paired inputs, with CogVideo-X, a transformer-based video diffusion model. To enhance spatial consistency and guide the generative process, we introduce an observable mask as an additional condition for the diffusion model. We further propose an iterative inference strategy that enables efficient generation of long sequences while maintaining high visual fidelity and multi-view consistency. Experiments on RealEstate-10K and ACID show that NoPo3DFusion sets a new state of the art for feed-forward NVS from sparse and unposed images, highlighting the effectiveness of combining pose-free 3D reconstruction with video diffusion model. Code and models are available at https://github.com/maxShen701/NoPo3DFusion .