Xuanyu Liu

3D indoor scene generation is starting to train design and critique as one capability. Earlier approaches either relied on general-purpose models without domain design experience or learned from static scenes, and raw construction trajectories carried much wasted trial and error. Scene-Designer learns construction and critique from two complementary datasets, then uses the learned critic to guide reinforcement learning over editing decisions; against the supervised-only Scene-Builder, the share of interaction-orientation errors needing five or more corrective moves falls from 40% to 7%, and window-accessibility blockages from 21% to 5%.

3D indoor scene generation is starting to train design and critique as one capability. Earlier approaches either relied on general-purpose models without domain design experience or learned from static scenes, and raw construction trajectories carried much wasted trial and error. Scene-Designer learns construction and critique from two complementary datasets, then uses the learned critic to guide reinforcement learning over editing decisions; against the supervised-only Scene-Builder, the share of interaction-orientation errors needing five or more corrective moves falls from 40% to 7%, and window-accessibility blockages from 21% to 5%.