WorldPager enables long interactive video world models on a single GPU
WorldPager: Zero-Stall Tiered Memory for Long-Horizon World Models on a Single Consumer GPU
Distributed, Parallel, and Cluster ComputingComputer Vision and Pattern Recognition
Summary
Interactive virtual world models keep track of past scenes to make revisiting places look consistent. Existing methods run out of memory or have short time limits on regular consumer GPUs. The authors show that memory size, not speed, limits how long these models can keep their memory. They create WorldPager, a system that stores and fetches memory efficiently across fast GPU memory, main memory, and SSD storage without delays. This lets the model generate much longer videos smoothly on one consumer GPU with minimal quality loss.
What this means in practice
- •For game developers: Create longer, more consistent virtual worlds that run on a single high-end gaming GPU without memory stalls or quality loss.
- •For augmented reality developers: Deliver extended, stable interactive AR experiences by managing large scene memories efficiently on consumer hardware.
Authors
Bingnan Liu, Ruini Xue
Abstract
Interactive world models keep the KV of past chunks as retrieval memory so that revisited places stay consistent. On a single consumer machine (one RTX 4090, 32 GiB of host memory, one NVMe SSD), long sessions with such memory hit two walls. The first is capacity: the KV of Matrix-Game 2.0 exhausts host memory after 45 seconds of video in bf16 and after 2.8 minutes in int4. The second is the model's positional horizon of 4.3 minutes. We show that this memory is capacity-bound rather than bandwidth-bound and that its access set is known before compute, because the camera pose is an input. An analytical model built on these observations predicts spill points to within one chunk and classifies zero versus non-zero stalls correctly on held-out configurations. Guided by it, WorldPager stores int4 KV in a GPU/DRAM/NVMe hierarchy instead of recomputing it, prefetches each chunk's memory from the extrapolated pose, never waits on a miss, and writes back in the background; to our knowledge, it is the first to place the KV memory of an interactive world model on an NVMe tier with zero-stall prefetching. With training-free position handling, it generates 37.5 minutes of video continuously with zero stalls and constant host memory, at about 9.5 frames per second. On a sealed test set, int4 storage shows no measurable quality loss; on long-gap revisits it lowers image quality by about one MUSIQ point.