FB-4D: Spatial-Temporal Coherent Dynamic 3D Content Generation with Feature Banks

• DOI: 10.48550/arxiv.2503.20784
• Publication Date: 2025-03-26
• Authors: Jinwei Li; Huan-ang Gao; Wenyi Li; Heng Chi; Chenyu Liu; Chunshui Du; Yiqian Liu; Mingju Gao; Guiyu Zhang; Zongzheng Zhang; Yi Li; Yao Yao; Jingwei Zhao; Hongyang Li; Yikai Wang; Hao Zhao
• Keywords: Feature (linguistics); Content (measure theory)

With the rapid advancements in diffusion models and 3D generation techniques, dynamic content has become a crucial research area. However, achieving high-fidelity 4D (dynamic 3D) with strong spatial-temporal consistency remains challenging task. Inspired by recent findings that pretrained features capture rich correspondences, we propose FB-4D, novel framework integrates Feature Bank mechanism to enhance both spatial temporal generated frames. In store extracted from previous frames fuse them into process of generating subsequent frames, ensuring consistent characteristics across time multiple views. To ensure compact representation, is updated proposed merging mechanism. Leveraging this Bank, demonstrate for first additional reference sequences through autoregressive iterations can continuously improve performance. Experimental results show FB-4D significantly outperforms existing methods terms rendering quality, consistency, robustness. It surpasses all multi-view tuning-free approaches large margin achieves performance on par training-based methods.