Visko 公布 Orbis 1.0 的评测结果：在自动化基准中于 DOVER 美学与技术和 VideoAlign 视觉与运动质量得分最高，并在 VideoPhy-2、Physics-IQ、VBench-2.0 Physics 三个物理合理性协议领先；在随机化人类 Arena 研究中总偏好和时间稳定性均排第一。
Most AI video benchmarks still reward how good a clip looks.
Orbis 1.0 from @viskoai is trying to prove something harder: whether a generated world can keep running without its motion, physics, or visual state falling apart.
In Visko's long-form human Arena, Orbis ranked first overall at 1,838 Elo and first in temporal stability at 1,940, ahead of the next systems at 1,734 and 1,807 respectively.
Orbis 1.0, a "Live Model" that keeps generating video continuously in real time, lets you change the prompt while the video is already running, and carries visual memory forward so the scene can persist for hour-scale sessions instead of ending after a short clip.
The unusual part is the combination: Visko says Orbis can stream 4K video at 24 FPS, react visibly to new instructions in under a second on average, and maintain stronger long-duration stability than the other real-time systems it tested, making it closer to a continuously running interactive visual world than a conventional text-to-video generator.
Our evaluation results for Orbis 1.0: Across automated benchmarks, Orbis scores highest on DOVER aesthetic and technical quality, and on VideoAlign visual and motion quality. It also leads three physical-plausibility protocols: VideoPhy-2, Physics-IQ, and VBench-2.0 Physics. In a randomized human Arena study of state-of-the-art real-time interactive video systems, Orbis ranked first in overall preference and first in temporal stability, the prerequisite for long-form generation. Full report: https://www.visko.ai/blog/papers/files/Visko_Orbis_1_0__A_Live_Model_for_Real_TimeInteractive_Long_Video_Generation.pdf在 X 查看被引用的帖子
来源：Rohan Paul · x.com