Video world model stores scene history in viewpoint-specific tokens

single source· 2 articles · confidence: low · first seen 2026-09-09 20:00 UTC

What this means for you

Nothing to act on: no code, no weights, no demo and no numbers, so the consistency claim cannot be checked yet. If you work on video generation, the idea worth watching is holding scene history in a fixed set of viewpoint-specific tokens instead of matching frames by depth.

WorldCrafter, posted to arXiv on 20 September, is a video world model — a system that generates a scene a camera can move through — that keeps a 3D-aware memory of what it has already produced. Rather than matching earlier frames by depth, it packs multi-view history into a fixed set of tokens chosen for each viewpoint, trained jointly with the video generator and sped up by distillation (a small model trained to imitate a larger one). The paper claims better long-horizon consistency and camera control over minute-scale exploration, with no numbers, no evaluation date and no code. A second, unrelated world-model paper from 9 September arrived without text.

Key facts

  • ·WorldCrafter is described as a video world model with a camera-queryable implicit 3D-aware memory. source
  • ·Its memory encoder and pose-conditioned readout module are trained jointly with the video generator, without explicit depth-based correspondences. source
  • ·The system generates streaming scene exploration from a single input image or a text prompt. source
  • ·The paper reports gains in long-horizon consistency and camera-control accuracy during minute-scale exploration, but gives no numeric results and no evaluation date. source
  • ·WorldCrafter was posted to arXiv on 20 September 2026 as arXiv:2609.24984. source
  • ·A separate world-model paper, arXiv:2609.11548, was posted on 9 September 2026 with no body text available. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire