Tsinghua's PPMem stores AI video memory as flat planes

A Tsinghua University-led collaboration introduced PPMem, a method for preserving memory in generative video world models. Instead of volumetric 3D point clouds, it stores flat surfaces such as walls and floors as 7-parameter plane primitives, keeping previously seen scenes consistent when the camera returns.

• Each plane primitive is defined by 7 numbers: center depth, two extents and a rotation quaternion
• The approach replaces millions of 3D points with reusable geometric planes
• Stored planes are projected back into view to guide generation on revisits
• It builds on Nvidia's HyDRA, Oxford's VMEM and Stanford's SPMem