WORLDMEM: Long-term Consistent World Simulation with Memory
Zeqi Xiao, Yushi Lan, Yifan Zhou, Wenqi Ouyang, Shuai Yang, Yanhong Zeng, Xingang Pan
2025-04-18
Summary
This paper talks about WorldMem, a new system that helps computers create and remember 3D worlds in a way that stays consistent over time and from different viewpoints, by using a special kind of memory.
What's the problem?
The problem is that when computers try to simulate or generate 3D scenes, they often have trouble keeping everything consistent, especially if the scene changes over time or if you look at it from different angles. This can make the virtual world look unrealistic or confusing.
What's the solution?
The researchers built WorldMem, which uses a memory bank along with an attention mechanism. This setup allows the computer to store important details about the 3D world and refer back to them as it generates new scenes or updates existing ones, making sure everything fits together smoothly no matter how the scene changes or where you look from.
Why it matters?
This matters because it makes virtual worlds in games, simulations, and virtual reality much more realistic and believable. It also helps with training robots or AI that need to understand and remember complex environments over time.
Abstract
WorldMem uses a memory bank with attention mechanism to enhance scene generation by preserving 3D spatial and temporal consistency across viewpoints and time.