Peking University Shenzhen Graduate School-RabbitPre AIGC Joint Research Laboratory has introduced the Open-Sora Plan, which aims to reproduce OpenAi’s Sora. People around the world are encouraged to join the open-source community and use a shared model to generate videos from text descriptions.

The model is part of the Tuling big model, an image and video generating AI platform, which was developed by RabbitPre and deeply involved Peking University Shenzhen Graduate School. The Tuling big model has now been applied in the industrial manufacturing, architecture, and fashion industries.

Since launching March 1 on GitHub, a global platform for software development, collaboration, and security, the Open-Sora Plan has garnered significant attention. It has also attracted participants from countries around the world, such as Australia, the U.K., Dubai, Iran, and Saudi Arabia.

“They have joined, used the model, and made contributions to it,” said Dong Shaoling, founder of RabbitPre, during a seminar held to introduce the Open-Sora Plan yesterday.

Between March 4 and 11, the project recorded more than 140,000 views on GitHub and has been cloned over 800 times. The model and code have been updated more than 200 times through the joint efforts of those who joined the plan, Dong said.

The plan has mainly attracted tech developers and people in relevant industries where the model can be applied, such as architecture and industrial manufacturing.

Those who join the plan are free to use the model and apply it to real-world scenarios, according to Yuan Li, deputy director of the joint research lab.

“The open community also provided us with positive feedback,” Yuan said. There are brilliant algorithm engineers and scientists who could help optimize the model, and companies and institutes could support the model with computing power or provide data to enrich its knowledge pool, Yuan added.

Yuan outlined the plan’s three phases. They have to determine whether the plan is practical in the initial phase, which runs until March 30. Based on current feedback, it appears feasible.

Their goal for the second phase is to train the model to generate videos that are over 20 seconds in length and achieve 720P video quality.

For phase three, they aim to match Sora’s ability to generate high-quality videos.