HiDream.ai launched an interactive world model that tops the WBench navigation board
HiDream-O1-World, announced August 24, generates explorable 3D worlds from a text prompt, image, or interactive control, and posted an average 80.9 on the Navi sub-leaderboard of WBench, the interactive world model benchmark built by Meituan's LongCat team with Fudan University, plus 73.3 on Physical and 88.0 on Consistency across 289 multi-turn cases and 1,058 interaction turns. The technical basis is 3D priors injected into a memory context plus Test-Time Training for online adaptation during inference, which is what the company credits for eliminating scene drift across viewpoint switches. The underlying paper, DreamWorld, is confirmed accepted to ECCV 2026 with authors from Fudan and HiDream.ai, though code on the project page is still marked coming soon and the launch numbers come from a company press release.
↳ Follow the thread