Open-Sora Plan is a community-driven research project led by Peking University’s YuanGroup that aims to reproduce OpenAI’s Sora text-to-video model in the open. It releases the architecture, training code and model checkpoints needed to generate videos from text prompts, and it has collected about 12,200 stars on GitHub.
When Sora demonstrated long, coherent text-to-video generation behind closed doors, Open-Sora Plan set out to rebuild the recipe publicly. Its pipeline combines a wavelet-flow video autoencoder (WF-VAE) that compresses video efficiently, a spatial-temporal diffusion transformer (ST-DiT) that generates the footage, and multilingual text conditioning, with the full training and inference stack released as open code. Successive versions have steadily improved resolution, duration and motion quality, and the project publishes detailed technical reports so others can learn from โ and build on โ each step. Contributions come from universities and companies across the open-source community, making the repository a shared lab notebook for open video generation. It is a research project rather than a polished consumer product: expect to bring your own GPUs and read the reports, but in return you get one of the most transparent windows into how modern text-to-video systems actually work.
Open-Sora Plan is free and open source under the MIT license. Models and code cost nothing; the real investment is the GPU compute needed to train or run video diffusion models.
Open-Sora Plan is for researchers, students and ML engineers who want to study, reproduce or extend state-of-the-art text-to-video generation with fully open code โ and for teams evaluating open video models as a base for their own products.