HiDream's world model scores 80.9 on navigation benchmark

HiDream built a model that turns a text prompt, an image, or interactive controls into an explorable 3D world. That matters for anyone building game prototypes, robotics simulations, or synthetic training data. A world model that can't hold a scene together when you look away twice isn't usable for a walkthrough demo. It's useless in a simulation loop that runs thousands of steps.

HiDream-O1-World scored 80.9 average on WBench's Navi benchmark, 73.3 on Physical, and 88.0 on Consistency, across 289 multi-turn test cases, per DreamWorld's project page. Consistency is the score to watch. Generative world models have a known failure mode where the walls move if you turn the camera away and back.

HiDream says it fixes this by injecting 3D priors into the model's memory context. It also runs test-time training so the model adapts online instead of relying only on what it learned during training. The company credits that combination for eliminating scene drift across viewpoint switches.

The underlying paper is accepted to ECCV 2026. Code is marked coming soon, so there's no way yet to verify the 80.9 and 88.0 figures independently. The launch numbers also come from HiDream's own press release, not a third-party benchmark run.

The next checkpoint is the code release. Until outside teams can run WBench against HiDream-O1-World themselves, the consistency claim is the company's word against a benchmark it didn't design.