Guide · Open source
How to use Open-Sora-Plan: setup, models and what to expect
Updated · written with AI from the project's own documentation and checked against our catalogue data
Open-Sora-Plan is an open research project from PKU-YuanGroup that aims to reproduce OpenAI's closed Sora text-to-video model. This guide explains what the repository contains, how to get the code, where the published model files are listed, and whether it fits your work. By the end you will know the first setup step and which parts of the project are meant for researchers, engineers and contributors.
Open-Sora PlanPKU-YuanGroup/Open-Sora-Plan · ★ 12.2k · MIT Install options & downloads Quick start
- 1Decide if this project fits your goal. Choose it if you want to study or contribute to open video generation models, not if you only want quick videos.
- 2Check your hardware. Confirm whether you have Huawei Ascend 910-series chips for v1.5.0, or enough VRAM for an older version as described in the README.
- 3Clone the repository. Run git clone https://github.com/PKU-YuanGroup/Open-Sora-Plan in your chosen folder.
- 4Pick a version. Compare the versions in the resource table and their reports to choose the one that matches your hardware and interest.
- 5Download the matching model files. Use the Hugging Face links in the resource table for the version you chose.
- 6Read the technical report. Open the report for your version in the docs folder to learn the architecture and how the model was trained.
What Open-Sora-Plan is and who it is for
Open-Sora-Plan is a research repository whose stated goal is to reproduce Sora, OpenAI's text-to-video model, in an open way. The project publishes code, model checkpoints and technical reports for its versions. The README says the team welcomes contributions from the open source community.
It is aimed at people who want to study or build video generation models. It is not a simple end-user app for making videos quickly.
- Researchers studying open video diffusion models and their architecture
- Engineers with Huawei Ascend hardware who want to run the v1.5.0 code
- Contributors who want to join an open development effort on text-to-video models
Clone the repository
The README's only install instruction is to clone the repository with Git. Run the command below in the folder where you want the project to live. The README lists no other prerequisites, so check the repository files for setup details before you build anything.
git clone https://github.com/PKU-YuanGroup/Open-Sora-PlanCheck your hardware before you start
The README does not list general hardware requirements. It does name two points you should know about. The v1.5.0 version is fully trained and inferred on Ascend 910-series accelerators, and the GPU version is described as coming soon. For an older release, the README says version 1.3.0 can run 93x480p video within 24G VRAM.
If you do not have Ascend hardware, the v1.5.0 README guidance does not cover your setup, so plan to read the repository documentation closely first.
- v1.5.0: trained and run on Huawei Ascend 910-series chips, per the README
- v1.3.0: the README mentions 93x480p within 24G VRAM
- GPU support for the newest version: described as coming soon
Find the model checkpoints and reports
The resource table links the model files on Hugging Face. For v1.5.0 it lists a diffusion model at 121x576x1024 and a CausalVideoVAE named Anysize_8x8x8_32dim, both under the MindSpeed folder. For v1.3.0 it lists Anysize diffusion models at 93x640x640, including an image-to-video variant, plus a VAE and a prompt refiner.
Technical reports are linked in the news section, including Report-v1.5.0.md and the older v1.3.0, v1.2.0, v1.1.0, v1.0.0 reports in the docs folder. Read the report for the version you pick, since each one describes different features.
- v1.5.0 uses an SUV (Skiparse 3D) architecture, per the resource table
- v1.5.0 code is on the mindspeed_mmdit branch, according to the news section
- v1.3.0 includes WFVAE, a prompt refiner, data filtering, sparse attention and bucket training, per the news section
Know what the project includes
The highlights describe a WFVAE with an 8×8×8 downsampling rate and a sparse DiT called SUV. The README claims SUV reaches performance close to dense DiT with over a 35% speedup. The news section says v1.5.0 is comparable to HunyuanVideo (Open-Source) using an 8B-scale model and 40 million video samples.
These are claims from the project's own README and reports. Check the linked technical report before relying on them in your own work.
What people use it for
- Studying the WFVAE video autoencoder and how it compresses video for a diffusion transformer
- Running the v1.5.0 code on Huawei Ascend hardware for video generation research
- Comparing the SUV sparse attention design against dense diffusion transformers
- Contributing code or experiments to an open text-to-video project
Frequently asked questions
Is Open-Sora-Plan free to use?
The repository is listed under the MIT licence. The README badge in the source names Apache, so check the LICENSE file in the repository itself before you rely on either.
What are the requirements to run Open-Sora-Plan?
The README does not list general requirements or prerequisites. It mentions Huawei Ascend 910-series chips for v1.5.0 and 24G VRAM for an older version at 93x480p.
Can I run Open-Sora-Plan on an NVIDIA GPU?
The README says the GPU version is coming soon for the newest release. Older versions are described in the README, but check the repository for GPU-specific instructions.
Is Open-Sora-Plan the same as Open-Sora?
No. Open-Sora-Plan is a separate PKU-YuanGroup project. Open-Sora is one of the comparable projects listed for it, so compare their reports and code before choosing.
How does Open-Sora-Plan compare with CogVideo, HunyuanVideo or Wan2.1?
The README compares v1.5.0 with HunyuanVideo and says its WFVAE achieves higher PSNR than the VAE used in Wan2.1. It does not give a full benchmark against the others, so review each project's own report.
Is there a simple app to generate videos with Open-Sora-Plan?
The README describes setup and model files, not an end-user app. If you want to make videos quickly, pick a tool built for that purpose.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Open-Sora-Plan suits researchers, engineers with Ascend hardware and contributors who want to study or extend an open text-to-video model. It is a poor fit if you want a polished, production-ready tool or a simple way to generate videos, since the README describes rapid iteration and a GPU version that is still coming soon.
Install options & downloads