On August 17, HiDream.ai, an artificial intelligence company based in Hefei, officially launched HiDream-O1-World, a native full-modality interactive world model. Based on the company’s self-developed UiT native full-modality architecture, the model integrates three core functions: navigation, editing and interaction. It supports multimodal inputs such as text, images and interactive commands, can generate a continuously explorable and interactive digital world with one click, and achieves breakthroughs in two key capabilities: spatiotemporal consistency and physical consistency.

Unlike traditional video generation models that merely produce continuous frames, interactive world models are required to sustain scene stability during users’ continuous perspective shifting, environment modification and character manipulation, while ensuring that object movements comply with the basic laws of the real world.

To address this challenge, HiDream-O1-World achieves significant improvements in long-horizon spatiotemporal consistency and physical consistency. By adopting 3D spatial memory and test-time training mechanisms, the model continuously records the geometric structure, object positions and spatial relationships of scenes during generation. This maintains relatively stable object sizes, positions and occlusion relationships after multiple camera movements and perspective switches, mitigating issues including scene drift and object disappearance.

Source: Hefei Evening News

Translated by Shao Hanzhe, intern