nv-tlabs/ATISS

Generates diverse 3D indoor scenes from floor plans and room layouts using an autoregressive transformer that sequentially predicts object categories, sizes, and poses. Built on PyTorch with fast-transformers, it trains on 3D-FRONT and 3D-FUTURE datasets and includes preprocessing pipelines for extracting scene geometry, object annotations, and top-down renderings. Provides pretrained models, visualization tools via simple-3dviz, and supports interactive scene editing tasks described in the paper.

330 stars. No commits in the last 6 months.