Original title: OctWorld: Long-Range World-Consistent Video Generation with Octree-Based 3D Mapping
Zelong Lv, Sicheng Xu, Jianfeng Xiang, Ruicheng Wang +4 · September 2026
When computers generate videos of large 3D spaces, keeping the world consistent is hard. As a virtual camera moves along long paths, places it already showed often distort or change when visited again. To solve this, researchers built OctWorld. Starting from a single picture, it creates high-quality, explorable scenes along a path chosen by a user. It works using a dynamic memory system called OctMap. OctMap takes new images and their depth maps, which show how far away surfaces are, and merges them into a global 3D tree. The tree adjusts its detail to match the scene without using too much memory. OctWorld produces long, spatially steady videos, beating previous methods and outperforming older point-based and fixed-resolution memory designs.
Imagine you are drawing a large map of a park as you walk through it, but you can only see a few feet ahead at a time. If you do not remember what you already drew, when you loop back, the paths and trees will not line up. Here, the system generates a video by moving through a scene, and OctMap acts like that memory, constantly updating a 3D sketch so the world stays consistent even when you revisit spots.
For anyone creating virtual worlds, like game designers or VR artists, this makes it possible to generate long, explorable scenes from a single image without the world falling apart when you revisit places.
The abstract does not state specific limitations, but the work focuses on long-range generation with extended camera paths, so its performance in other settings like short videos or real-time applications is not established.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
OctWorld: Long-Range World-Consistent Video Generation with Octree-Based 3D Mapping — Zelong Lv, Sicheng Xu, Jianfeng Xiang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.