OctWorld presented as a generator of long consistent videos with octree-based 3D memory

On September 3, 2026, the paper OctWorld was published on arXiv — a framework for video generation based on diffusion with persistent 3D memory. From a single image, it builds long video sequences along a given camera trajectory, maintaining scene consistency even when returning to previously visited areas. The key element is OctMap, a sparse octree with TSDF fusion that accumulates visual observations and depth maps. The authors claim that OctMap outperforms point caches and TSDF volumes with fixed resolution, and they have not released the code. Details are on the paper page on arXiv.

What it means

OctWorld solves the task of video generation, not 3D model creation, so it does not directly compete with the tools in our section of image-to-3D generators. However, a comparison in terms of memory and approach to 3D representation is appropriate. OpenLRM has the largest model weights file at 1.7 GB, all files total 3.4 GB, the task is image-to-3d, with export to obj. It is a generator of static models, not video sequences, and does not need 3D memory in the spirit of OctMap. Hunyuan3D 2.1 is heavier: the largest model weights file is 6.9 GB, all files total 13.9 GB, and it requires 10 to 29 GB of VRAM. It also creates static 3D assets from a single image, without camera trajectories or long-term video consistency. TRELLIS is a 3D model generator with the largest model weights file at 1.1 GB and export to glb. Like the others, it does not work with video sequences and does not store 3D memory between frames. OctWorld is interesting because it brings the idea of persistent 3D representation into video generation. For users of our tools, this is still more of a research work: there is no code, and the task is long consistent videos, not 3D assets.