wu-cvgl/neuworld

NeuWorld is the code repository for a research paper called Walking in the Implicit: Interactive World Exploration via Neural Scene Representation, accepted at the ECCV 2026 computer vision conference. The paper comes from researchers at Zhejiang University, Westlake University, and a company called Afari Intelligent Drive. As of this README, no code has been published yet. The repository states plainly that the public code is under internal review and that code and model checkpoints will be released soon. Right now the repository mainly hosts the project's branding, a paper PDF, links to the arXiv preprint and project page, and a placeholder link for Hugging Face model checkpoints. The research is about letting a computer generate and explore a scene interactively, similar to walking around inside a space, rather than only producing a straight video. Instead of generating a growing sequence of video frames, the method keeps a fixed-size internal representation of the scene, called a Neural Implicit Scene state, and updates it at each step. The system separates two things at every step: a random update to the internal scene representation, and a separate, predictable step that renders what a camera would see from a given position. Camera movement, a reference image, and relevant pieces of past scene history are all fed into the model through the same shared representation rather than separate specialized parts. The authors say this design lets a user explore a scene over a long stretch of time, including returning to a place already visited, while keeping the system reasonably fast to run. The model is trained entirely from scratch on public datasets of images with known camera positions, without relying on existing pretrained video generation models or separate 3D reconstruction tools. The README does not state a license for the eventual code release.