Qualitative comparison with 4D control baselines. Compared with FlexTraj and VerseCrafter, VoxelMem better preserves foreground identity and background layout when the camera moves away and returns, benefiting from voxel-based local and global memory.

Qualitative comparison with long video generation baselines. MG-3 lacks precise camera and object control, while Spmem and LWorld struggle with dynamic foregrounds. VoxelMem better preserves scene structure and foreground identity through unified 4D voxel control and memory.

Ablation study on VoxelMem. Removing local or global memory weakens appearance preservation and revisit consistency.

Ablation study on VoxelMem. Channel-Only injection or latent-patch warping introduces artifacts. The full hybrid memory design achieves the most stable long video generation.

Limitations in Handling Voxel Splitting and Merging. VoxelMem may fail to faithfully follow prescribed fragmentation during voxel splitting or propagate appearance changes after voxel merging, such as transferring the red color of a droplet into the pool.