EVENT DOSSIER
AnchorWeave is a memory-enhanced video generation framework designed to improve consistency in long-term scenarios. This approach abandons a single global memory and instead uses multiple clean local geometric memories to replace the error-prone misaligned global memory. The system integrates the locally retrieved memories for target trajectory alignment through a coverage-driven retrieval mechanism and a multi-anchor weaving controller, thereby addressing inconsistencies across views. Experimental results show that AnchorWeave significantly improves consistency in long-term scenarios while maintaining high visual quality. A ablation study further confirms the effectiveness of local geometric conditions, multi-anchor control, and coverage-driven retrieval strategies.
AnchorWeave 提出基于局部几何记忆的视频生成框架
AnchorWeave 提出一种记忆增强视频生成框架，以多个清洁局部几何记忆替代单一错位全局记忆，通过覆盖驱动的检索与多锚编织控制器解决跨视图不一致问题。该方法利用目标轨迹对齐检索局部记忆，并在生成过程中整合这些记忆。实验表明，AnchorWeave 显著提升了长时场景一致性并保持高质量视觉表现，消融研究进一步验证了局部几何条件、多锚控制及覆盖驱动检索的有效性。