NuiScene
NuiScene is a research software introduced at ICCV 2025 for the efficient generation of unbounded outdoor 3D scenes. Developed by Han-Hung Lee, Qinghong Han, and Angel X. Chang, it employs a two-stage pipeline involving a variational autoencoder to compress scene data into learned embeddings and a diffusion model trained on these compressed representations to generate large-scale environments. The system supports infinite or unbounded scene synthesis by breaking the generation process into manageable chunks, allowing users to define the total size of the output scene through configuration parameters. It includes pretrained models for single-scene, four-scene, and thirteen-scene datasets, with flexibility for generating custom grid sizes. The software outputs 3D geometry in OBJ format for visualization and further processing. Key optimizations include anti-diagonal batch processing to accelerate inference and advanced vector set upsampling for improved reconstruction quality in multi-scene models. NuiScene is