Scene Dreamer

SceneDreamer is a research project from Nanyang Technological University, published in IEEE Transactions on Pattern Analysis and Machine Intelligence in 2023, that generates large, unbounded 3D landscapes. The model is trained only on collections of ordinary 2D photos, without 3D annotations. Starting from random simplex noise, it builds a bird's-eye-view representation made of a height field for terrain elevation and a semantic field for scene content, encodes features with a generative neural hash grid, and renders photorealistic views with a style-modulated volumetric renderer trained adversarially. The result is diverse landscapes in different styles with consistent depth that a virtual camera can fly through freely.

SceneDreamer is aimed at researchers and developers in 3D generation, graphics and world-building who want to study or reproduce the method. It is academic code rather than a supported product, so setup, hardware needs and output quality depend on the research release. Nothing was tested for this listing.

There is no pricing: the project page links to the paper, a public code repository and an online demo. The software licence was not reviewed for this listing, so users should check the repository's terms before reuse.