OpenSora
Fully open-source text-to-video generation — 11B model, training code included.
About
OpenSora is HPC-AI Tech's initiative to democratize Sora-class video generation: model weights, training code, data preprocessing, and technical reports are all public under Apache 2.0. Open-Sora 2.0 (11B parameters) reports performance on par with HunyuanVideo, generating up to 15-second 720p clips across text-to-video, image-to-video, and text-to-image tasks via its STDiT spatial-temporal diffusion transformer.
With ~30k GitHub stars it is one of the most-watched open video projects, and the team also runs Video Ocean as a hosted product on the same tech. Be realistic about hardware: training is a $200K-scale endeavor and even inference wants serious GPUs — this is a research and builder's toolkit, not a consumer app.
What it can do
- Text-to-video and image-to-video generation
- 11B open model with full training pipeline
- Up to 15s 720p output, arbitrary aspect ratios
- Gradio demo and inference scripts
Setup
Clone https://github.com/hpcaitech/Open-Sora, install dependencies, download Open-Sora-v2 weights from Hugging Face, run the Gradio demo or inference scripts.
Prerequisites
Python 3.10, multi-GPU setup recommended, Hugging Face or ModelScope access for weights.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 63; Trending score 82/100 from researcher assessment.
★ 29,871 GitHub stars
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.