Chinese World Model Tops Li Feifei's Ranking: Adapted for Ascend Compute, Full Code and Weights Open-Sourced
Shanghai Artificial Intelligence Laboratory
The Shanghai AI Laboratory team unveiled OpenEM, a world model that ranked first in Li Feifei's VBench leaderboard. The model is adapted for Chinese Ascend chips and is fully open-sourced (code, weights, data), enabling 3D scene generation and interaction.
The Shanghai AI Laboratory team has unveiled the OpenEM world model, capable of generating a 3D scene from a single image, offering freedom of movement, interactive editing, and weather control. The model ranked first in the VBench leaderboard (Fei-Fei Li's team), scoring 85.75 points, surpassing CogVideoX. OpenEM is adapted for Chinese Ascend 910B chips (Huawei), uses 3D Gaussian representation, a video diffusion model, and 6D video Parquet for stable generation of up to 300 frames. All code, weights, and data are open-sourced under the MIT license.
Source: QbitAI 量子位 —
original