Kunming Luo, Ping Tan et al. · 3D Scene Generation

DistScene

Generates complete 3D scenes from images, producing the environment and the objects in it together. Its training data is made automatically: TRELLIS.2 (covered here in week 51 of 2025) creates 3D objects and environments, physics checks guide how they are placed, and rendered views become the input images. The output scenes load directly into navigation simulators. Paper arXiv 2610.06960 (3 October) by 11 authors including Kunming Luo, Xuelong Li and Ping Tan. Preview assets are on Hugging Face; code, checkpoint and dataset are listed as coming soon.