Memory module也可以认为是History module
为什么要设计Memory module：
Memory module的设计范式：
范式1：3D-based Memory
Memory：构建3D surfels，每个surfel记录past views。
怎么写入Memory：常规的surfel reconstruction方法，在每个surfel中记录相应的view。
怎么读取Memory：基于novel view渲染surfels，统计visible surfers中的past views，选取top-k最多可见的past views。
2025.06，VMem: Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory
Memory：构建3D point clouds。
怎么写入Memory：常规的点云重建方法。
怎么读取Memory：渲染点云到novel view，作为condition。
2025.06，Video World Models with Long-term Spatial Memory
范式2：Token-based Memory
Memory：Long and short-term memory，既包含了短期的high-resolution tokens, 又包含了长期的coarse tokens。
怎么写入Memory：存储generated tokens，并做一些处理。
怎么读取Memory：根据Routing机制读取。
范式3：Image-based Memory
Memory：Generated frames。
怎么写入Memory：存储生成的图片。
怎么读取Memory：根据camera pose选取Top-K最接近的history images作为condition。
2025.04，Long-term Consistent World Simulation with Memory
2025.05，Learning World Models for Interactive Video Generation
2025.06，Context as Memory: Scene-Consistent Interactive Long Video Generation with Memory Retrieval
2025.06，WorldExplorer: Towards Generating Fully Navigable 3D Scenes
2025.06，DeepVerse: 4D Autoregressive Video Generation as a World Model
范式4：Network-based Memory
2025.03，WorldWeaver: Generating Long-Horizon Video Worlds via Rich Perception
2025.04，WORLDMEM: Long-term Consistent World Simulation with Memory
2025.05，Learning World Models for Interactive Video Generation
2025.06，Video World Models with Long-term Spatial Memory
2025.06，WorldExplorer: Towards Generating Fully Navigable 3D Scenes
2025.06，DeepVerse: 4D Autoregressive Video Generation as a World Model
2025.06，Hunyuan-GameCraft: High-dynamic Interactive Game Video Generation with Hybrid History Condition